{"id":25389,"date":"2025-04-11T09:55:08","date_gmt":"2025-04-11T01:55:08","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=25389"},"modified":"2025-04-11T09:55:08","modified_gmt":"2025-04-11T01:55:08","slug":"%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%8e%a8%e7%90%86%e6%97%a0%e6%8d%9f%e5%8a%a0%e9%80%9f6-5%e5%80%8d%ef%bc%81eagle-3%e7%a2%be%e5%8e%8b%e4%b8%80%e5%88%87%e3%80%81%e5%bb%b6%e7%bb%adscaling-law%e8%83%bd","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2025\/04\/11\/25389\/","title":{"rendered":"\u5927\u6a21\u578b\u63a8\u7406\u65e0\u635f\u52a0\u901f6.5\u500d\uff01EAGLE-3\u78be\u538b\u4e00\u5207\u3001\u5ef6\u7eedScaling Law\u80fd\u529b"},"content":{"rendered":"
\"\"<\/div>\n
<\/div>\n
\u8fd1\u65e5\uff0cEAGLE \u56e2\u961f\u7684\u65b0\u4f5c\u300aEAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test\u300b\u901a\u8fc7\u4e00\u7cfb\u5217\u4f18\u5316\u89e3\u9501\u4e86\u6295\u673a\u91c7\u6837\u7684 Scaling Law \u80fd\u529b\uff0c\u53ef\u4ee5\u5c06\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u901f\u5ea6\u63d0\u9ad8 6.5 \u500d\uff0c\u540c\u65f6\u4e0d\u6539\u53d8\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8f93\u51fa\u5206\u5e03\uff0c\u786e\u4fdd\u65e0\u635f\u3002<\/div>\n
\n
\n

\u81ea\u56de\u5f52\u89e3\u7801\u5df2\u7136\u6210\u4e3a\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u6807\u51c6\u3002\u5927\u8bed\u8a00\u6a21\u578b\u6bcf\u6b21\u524d\u5411\u8ba1\u7b97\u9700\u8981\u8bbf\u95ee\u5b83\u5168\u90e8\u7684\u53c2\u6570\uff0c\u4f46\u53ea\u80fd\u5f97\u5230\u4e00\u4e2a token\uff0c\u5bfc\u81f4\u5176\u751f\u6210\u6602\u8d35\u4e14\u7f13\u6162\u3002<\/p>\n

\u8fd1\u65e5\uff0cEAGLE \u56e2\u961f\u7684\u65b0\u4f5c\u300aEAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test\u300b\u901a\u8fc7\u4e00\u7cfb\u5217\u4f18\u5316\u89e3\u9501\u4e86\u6295\u673a\u91c7\u6837\u7684 Scaling Law \u80fd\u529b\uff0c\u53ef\u4ee5\u5c06\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u901f\u5ea6\u63d0\u9ad8 6.5 \u500d\uff0c\u540c\u65f6\u4e0d\u6539\u53d8\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8f93\u51fa\u5206\u5e03\uff0c\u786e\u4fdd\u65e0\u635f\u3002\u540c\u65f6\uff0c\u968f\u7740\u8bad\u7ec3\u6570\u636e\u7684\u589e\u52a0\uff0c\u52a0\u901f\u6bd4\u8fd8\u80fd\u8fdb\u4e00\u6b65\u63d0\u9ad8\u3002<\/p>\n

\"\u56fe\u7247\"<\/p>\n

    \n
  • \u8bba\u6587\u6807\u9898\uff1aEAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test<\/li>\n
  • arXiv \u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2503.01840<\/li>\n
  • \u9879\u76ee\u5730\u5740\uff1ahttps:\/\/github.com\/SafeAILab\/EAGLE<\/li>\n
  • SGLang \u7248\u672c\uff1ahttps:\/\/github.com\/sgl-project\/sglang\/pull\/4247<\/li>\n<\/ul>\n

    EAGLE-3 \u7684\u52a0\u901f\u6548\u679c\uff08DeepSeek-R1-Distill-LLaMA 8B \u5728\u6570\u5b66\u63a8\u7406\u6570\u636e\u96c6 GSM8K \u4e0a\u6d4b\u8bd5\uff0c\u5176\u4ed6\u6a21\u578b\u5728\u591a\u8f6e\u5bf9\u8bdd\u6570\u636e\u96c6 MT-bench \u4e0a\u6d4b\u8bd5\uff09\uff1a<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u4e0d\u540c\u65b9\u6cd5\u7684\u751f\u6210\u901f\u5ea6\u5bf9\u6bd4\uff1a<\/p>\n

    <\/p>\n

    \u80cc\u666f<\/h4>\n

    \u6295\u673a\u91c7\u6837\u4f7f\u7528\u4e00\u4e2a\u5c0f\u7684\u6a21\u578b\u5feb\u901f\u751f\u6210\u8349\u7a3f\uff0c\u4e00\u6b21\u751f\u6210\u591a\u4e2a token\u3002\u76ee\u6807\u5927\u8bed\u8a00\u6a21\u578b\u901a\u8fc7\u4e00\u6b21\u524d\u5411\u8ba1\u7b97\u5e76\u884c\u9a8c\u8bc1\u8349\u7a3f\u7684\u6b63\u786e\u6027\uff0c\u8f93\u51fa\u6b63\u786e\u7684\u8349\u7a3f\uff0c\u5e76\u786e\u4fdd\u65e0\u635f\u3002EAGLE \u7cfb\u5217\u662f\u6295\u673a\u91c7\u6837\u7684\u6700\u5feb\u5b9e\u73b0\u3002<\/p>\n

    EAGLE-1 \u5728\u66f4\u6709\u89c4\u5f8b\u7684\u7279\u5f81\u5c42\u9762\u800c\u4e0d\u662f token \u5c42\u9762\u8fdb\u884c\u81ea\u56de\u5f52\uff0c\u540c\u65f6\u8f93\u5165\u91c7\u6837\u7ed3\u679c\uff08\u8d85\u524d\u4e00\u4e2a\u65f6\u95f4\u6b65\u7684 token\uff09\u6d88\u9664\u4e86\u4e0d\u786e\u5b9a\u6027\uff0c\u660e\u663e\u63d0\u5347\u4e86\u8349\u7a3f\u6a21\u578b\u7684\u51c6\u786e\u7387\u3002EAGLE-2 \u5229\u7528\u8349\u7a3f\u6a21\u578b\u7684\u7f6e\u4fe1\u5ea6\u8fd1\u4f3c\u63a5\u53d7\u7387\uff0c\u636e\u6b64\u52a8\u6001\u5730\u8c03\u6574\u8349\u7a3f\u6811\u7684\u7ed3\u6784\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u4e86\u6295\u673a\u91c7\u6837\u7684\u6548\u7387\u3002<\/p>\n

    \u6700\u65b0\u7684\u5927\u6a21\u578b\u901a\u8fc7\u4f7f\u7528\u8d8a\u6765\u8d8a\u591a\u7684\u8bad\u7ec3\u6570\u636e\u4ee5\u53d6\u5f97\u66f4\u597d\u7684\u6027\u80fd\u3002\u6bd4\u5982\u8bf4\uff0c\u5bf9\u4e8e LLaMA \u7cfb\u5217 7B\uff088B\uff09\u5927\u5c0f\u7684\u6a21\u578b\uff0cLLaMA 1\u3001LLaMA 2 \u548c LLaMA 3 \u5206\u522b\u4f7f\u7528\u4e86 1T\u30012T\u300115T token \u8bad\u7ec3\u6570\u636e\uff0c\u6a21\u578b\u7ed3\u6784\u548c\u63a8\u7406\u6210\u672c\u57fa\u672c\u4e0d\u53d8\u7684\u524d\u63d0\u4e0b\u5404\u9879\u6307\u6807\u53d6\u5f97\u4e86\u660e\u663e\u63d0\u5347\u3002<\/p>\n

    EAGLE-3 \u7684\u4f5c\u8005\u4eec\u5e0c\u671b\u901a\u8fc7\u589e\u52a0\u8349\u7a3f\u6a21\u578b\u7684\u8bad\u7ec3\u6570\u636e\u91cf\u4ee5\u63d0\u9ad8\u63a5\u53d7\u7387\u548c\u52a0\u901f\u6bd4\uff08\u84dd\u8272\u66f2\u7ebf\uff09\u3002\u9057\u61be\u7684\u662f\uff0c EAGLE-1 \u4ece\u8bad\u7ec3\u6570\u636e\u589e\u52a0\u4e2d\u5f97\u5230\u7684\u63d0\u5347\u975e\u5e38\u6709\u9650\uff08\u7ea2\u8272\u66f2\u7ebf\uff09\u3002<\/p>\n

    EAGLE-1 \u548c EAGLE-3 \u7684\u52a0\u901f\u6bd4\u968f\u7740\u8bad\u7ec3\u6570\u636e\u7684\u589e\u52a0\u800c\u53d8\u5316\u7684\u8d8b\u52bf\uff1a<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u52a8\u673a<\/h4>\n

    EAGLE \u5728\u7279\u5f81\u5c42\u8fdb\u884c\u81ea\u56de\u5f52\uff0c\u518d\u4f7f\u7528\u76ee\u6807\u6a21\u578b\u7684\u5206\u7c7b\u5934\u5f97\u5230\u8349\u7a3f token\u3002\u8fd9\u79cd\u65b9\u5f0f\u5229\u7528\u4e86\u76ee\u6807\u6a21\u578b\u7684\u4e2d\u95f4\u7ed3\u679c\u548c\u4fe1\u606f\uff0c\u76f8\u6bd4 token \u5c42\u81ea\u56de\u5f52\u6709\u66f4\u597d\u7684\u6027\u80fd\u3002\u56e0\u4e3a\u8349\u7a3f\u6a21\u578b\u7684\u6700\u7ec8\u4efb\u52a1\u662f\u9884\u6d4b\u8349\u7a3f token\uff0cEAGLE \u7684\u635f\u5931\u51fd\u6570\u5305\u62ec\u4e24\u90e8\u5206\uff0c\u4e00\u90e8\u5206\u662f\u7279\u5f81\u9884\u6d4b\u635f\u5931\uff0c\u53e6\u4e00\u90e8\u5206\u662f token \u9884\u6d4b\u635f\u5931\u3002\u7279\u5f81\u9884\u6d4b\u635f\u5931\u4e5f\u53ef\u4ee5\u88ab\u8ba4\u4e3a\u662f\u4e00\u79cd\u7ea6\u675f\uff0c\u9650\u5236\u4e86\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\u3002<\/p>\n

    EAGLE-3 \u7684\u4f5c\u8005\u4eec\u8fdb\u884c\u4e86\u5b9e\u9a8c\u9a8c\u8bc1\uff0c\u89c2\u5bdf\u7b2c\u4e00\u4e2a\u8349\u7a3f token \u7684\u63a5\u53d7\u7387 0-\u03b1\uff0c\u5728\u6570\u636e\u96c6\u8f83\u5c0f\u65f6\uff0c\u7279\u5f81\u9884\u6d4b\u635f\u5931\u5728\u8bad\u7ec3\u6570\u636e\u8f83\u5c11\u65f6\u53ef\u4ee5\u63d0\u9ad8\u6027\u80fd\uff0c\u4f46\u4e25\u91cd\u635f\u5bb3\u4e86\u8349\u7a3f\u6a21\u578b\u7684 scaling up \u80fd\u529b\u3002<\/p>\n

    \"\u56fe\u7247\"\u4e0d\u8fc7\uff0c\u4e0d\u4f7f\u7528\u7279\u5f81\u9884\u6d4b\u635f\u5931\u4f1a\u5bfc\u81f4\u751f\u6210\u540e\u7eed\u8349\u7a3f token \u65f6\u7684\u8f93\u5165\u504f\u79bb\u8bad\u7ec3\u5206\u5e03\uff0c\u5bfc\u81f4\u7b2c\u4e8c\u4e2a\u8349\u7a3f token \u7684\u63a5\u53d7\u7387 1-\u03b1 \u4e25\u91cd\u4e0b\u964d\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0cEAGLE-3 \u4f7f\u7528 \u201c\u8bad\u7ec3\u65f6\u6d4b\u8bd5\u201d \u6a21\u62df\u591a\u6b65\u751f\u6210\uff0c\u4ece\u800c\u517c\u987e\u4e86\u6a21\u578b\u7684 scaling up \u80fd\u529b\u548c\u751f\u6210\u591a\u4e2a\u8349\u7a3f token \u7684\u80fd\u529b\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    EAGLE\u3001Medusa \u7b49\u6295\u673a\u91c7\u6837\u65b9\u6cd5\u90fd\u91cd\u7528\u76ee\u6807\u6a21\u578b\u7684\u6700\u540e\u4e00\u5c42\u7279\u5f81\u4f5c\u4e3a\u8349\u7a3f\u6a21\u578b\u7684\u63d0\u793a\uff0c\u4f46 EAGLE-3 \u7684\u4f5c\u8005\u4eec\u53d1\u73b0\u8fd9\u5b58\u5728\u7f3a\u9677\u3002\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6700\u540e\u4e00\u5c42\u7279\u5f81\u7ecf\u8fc7\u7ebf\u6027\u53d8\u6362\u5c31\u80fd\u5f97\u5230\u4e0b\u4e00\u4e2a token \u7684\u5206\u5e03\u3002\u5bf9\u4e8e\u6ee1\u79e9\u7684\u5206\u7c7b\u5934\uff0c\u6700\u540e\u4e00\u5c42\u7279\u5f81\u548c\u4e0b\u4e00\u4e2a token \u7684\u5206\u5e03\u4e00\u4e00\u5bf9\u5e94\u3002\u6700\u540e\u4e00\u5c42\u7279\u5f81\u53ea\u6709\u4e0b\u4e00\u4e2a token \u7684\u4fe1\u606f\uff0c\u5931\u53bb\u4e86\u76ee\u6807\u6a21\u578b\u7684\u5168\u5c40\u6027\u8d28\u3002<\/p>\n

    \u56e0\u6b64\uff0cEAGLE-3 \u4e0d\u518d\u4f7f\u7528\u76ee\u6807\u6a21\u578b\u7684\u6700\u540e\u4e00\u5c42\u7279\u5f81\u4f5c\u4e3a\u8f85\u52a9\u4fe1\u606f\uff0c\u800c\u662f\u6df7\u5408\u76ee\u6807\u6a21\u578b\u7684\u4f4e\u5c42\u3001\u4e2d\u5c42\u3001\u9ad8\u5c42\u4fe1\u606f\u6765\u4f5c\u4e3a\u8349\u7a3f\u6a21\u578b\u7684\u8f93\u5165\u3002<\/p>\n

    \u65b9\u6cd5<\/h4>\n

    \u4e0e\u5176\u4ed6\u6295\u673a\u91c7\u6837\u65b9\u6cd5\u4e00\u81f4\uff0cEAGLE-3 \u4ea4\u66ff\u6267\u884c\u8349\u7a3f\u9636\u6bb5\u548c\u9a8c\u8bc1\u9636\u6bb5\u3002<\/p>\n

    \u8349\u7a3f\u9636\u6bb5\u4ee5\u4e0b\u56fe\u4e3a\u4f8b\u3002\u5728\u9884\u586b\u5145\u6216\u8005\u4e0a\u4e00\u4e2a\u9a8c\u8bc1\u9636\u6bb5\uff0cEAGLE-3 \u8bb0\u5f55\u6a21\u578b\u7684\u4f4e\u3001\u4e2d\u3001\u9ad8\u5c42\u7279\u5f81\u5e8f\u5217\uff0c\u5206\u522b\u8bb0\u4e3a l\u3001m \u548c h\uff0c\u62fc\u63a5 k \u7ef4\u5411\u91cf l\u3001m \u548c h \u5f97\u5230 3k \u7ef4\u5411\u91cf\uff0c\u7136\u540e\u901a\u8fc7\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\u5c06\u5176\u964d\u7ef4\u5230 k \u7ef4\uff0c\u5c31\u5f97\u5230\u878d\u5408\u4e86\u4e0d\u540c\u5c42\u7ea7\u4fe1\u606f\u7684\u7279\u5f81 g\uff0c\u5176\u4e2d k \u4e3a\u76ee\u6807\u6a21\u578b\u7684\u9690\u85cf\u5c42\u7ef4\u5ea6\u3002\u76ee\u6807\u662f\u751f\u6210\u4ee5 \u201cHow can I\u201d \u4e3a\u524d\u7f00\u7684\u8349\u7a3f token \u5e8f\u5217\uff0c\u53ea\u8f93\u5165 g_how \u548c g_can\uff0c\u8349\u7a3f\u6a21\u578b\u65e0\u6cd5\u611f\u77e5\u968f\u673a\u7684\u91c7\u6837\u8fc7\u7a0b\u3002<\/p>\n

    \u56e0\u6b64 EAGLE-3 \u5f15\u5165\u91c7\u6837\u7ed3\u679c I \u7684\u8bcd\u5411\u91cf\u5d4c\u5165 e_I\u3002\u5c06 g \u4e0e\u5bf9\u5e94\u7684\u8d85\u524d\u4e00\u4e2a\u65f6\u95f4\u6b65\u7684\u5d4c\u5165 e \u62fc\u63a5\uff0c\u5373 g_how \u4e0e e_can \u62fc\u63a5\uff0cg_can \u548c e_I \u62fc\u63a5\u3002\u5c06\u62fc\u63a5\u540e\u7684\u5411\u91cf\u901a\u8fc7\u5168\u8fde\u63a5\u5c42\u964d\u5230 k \u7ef4\uff0c\u5e76\u5c06\u5176\u8f93\u5165\u8fdb\u4e00\u4e2a\u5355\u5c42\u7684 transformer \u5f97\u5230 a\u3002\u7136\u540e\uff0c\u5c06 a \u8f93\u5165\u5206\u7c7b\u5934\uff0c\u91c7\u6837\u5f97\u5230\u7b2c\u4e00\u4e2a\u8349\u7a3f token \u201cdo\u201d\u3002<\/p>\n

    \u5728 step 1\uff0c\u4ee5 \u201cHow can\u201d \u4e3a\u524d\u7f00\u65f6\uff0cEAGLE-3 \u91cd\u7528\u4e86\u6765\u81ea\u76ee\u6807\u6a21\u578b\u7684 g_how \u548c g_can\u3002<\/p>\n

    \u5728 step 2\uff0c\u524d\u7f00\u662f \u201cHow can I\u201d\u3002\u7406\u60f3\u7684\u65b9\u5f0f\u662f\u91cd\u7528\u6765\u81ea\u76ee\u6807\u6a21\u578b\u7684 g_how\u3001g_can \u548c g_I\u3002\u4f46\u8fd9\u662f\u4e0d\u53ef\u80fd\u7684\uff0c\u56e0\u4e3a token \u201cI\u201d \u8fd8\u6ca1\u6709\u88ab\u76ee\u6807\u6a21\u578b\u68c0\u67e5\uff0c\u65e0\u6cd5\u83b7\u53d6 g_I\u3002EAGLE-3 \u7528\u4e0a\u4e00\u4e2a step \u8349\u7a3f\u6a21\u578b\u7684\u8f93\u51fa a_I \u66ff\u4ee3 g_I\uff0c\u62fc\u63a5 a_I \u4e0e\u91c7\u6837\u7ed3\u679c \u201cdo\u201d \u7684\u8bcd\u5411\u91cf\u5d4c\u5165\u4f5c\u4e3a\u8349\u7a3f\u6a21\u578b step2 \u7684\u8f93\u5165\u3002<\/p>\n

    \u5728 step 3\uff0c\u540c\u6837\u65e0\u6cd5\u83b7\u5f97 g_do\uff0c\u6240\u4ee5\u4f7f\u7528 a_do \u4ee3\u66ff\uff0c\u62fc\u63a5 a_do \u4e0e e_it \u4f5c\u4e3a\u8349\u7a3f\u6a21\u578b\u7684\u8f93\u5165\u3002\u4e4b\u540e\u7684 step \u540c\u7406\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u5b9e\u9a8c<\/h4>\n

    EAGLE-3 \u5728\u591a\u8f6e\u5bf9\u8bdd\u3001\u4ee3\u7801\u3001\u6570\u5b66\u63a8\u7406\u3001\u6307\u4ee4\u9075\u5faa\u3001\u603b\u7ed3\u4e94\u9879\u4efb\u52a1\u4e0a\u5206\u522b\u4f7f\u7528 MT-bench\u3001Humaneval\u3001GSM8K\u3001Alpaca\u3001CNN\/DM \u6570\u636e\u96c6\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u5e76\u4e0e 7 \u79cd\u5148\u8fdb\u7684\u6295\u673a\u91c7\u6837\u65b9\u6cd5\uff08SpS\u3001PLD\u3001Medusa\u3001Lookahead\u3001Hydra\u3001EAGLE\u3001EAGLE-2\uff09\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002<\/p>\n

    \u8be5\u5b9e\u9a8c\u5206\u522b\u5728 Vicuna (V), LLaMA-Instruct 3.1 (L31), LLaMA-Instruct 3.3 (L33), DeepSeek-R1-Distill-LLaMA (DSL) \u4e0a\u8fdb\u884c\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u8868\u683c\u4e2d\u7684 Speedup \u4e3a\u52a0\u901f\u6bd4\uff0c\u03c4 \u4e3a\u5e73\u5747\u63a5\u53d7\u957f\u5ea6\uff0c\u4e5f\u5c31\u662f\u76ee\u6807\u6a21\u578b\u6bcf\u6b21\u524d\u5411\u8ba1\u7b97\u80fd\u751f\u6210\u7684 token \u6570\u3002EAGLE-3 \u6bcf\u6b21\u524d\u5411\u8ba1\u7b97\u80fd\u751f\u6210\u5927\u7ea6 4-7 \u4e2a token\uff0c\u800c\u81ea\u56de\u5f52\u89e3\u7801\u6bcf\u6b21\u751f\u6210 1 \u4e2a token\uff0c\u56e0\u6b64 EAGLE-3 \u660e\u663e\u52a0\u901f\u4e86\u5927\u8bed\u8a00\u6a21\u578b\u7684\u751f\u6210\uff0c\u52a0\u901f\u6bd4\u4e3a\u00a03.1x-6.5x\u3002<\/p>\n

    \u5728\u6240\u6709\u4efb\u52a1\u548c\u6a21\u578b\u4e0a\uff0cEAGLE-3 \u7684\u52a0\u901f\u6bd4\u548c\u5e73\u5747\u63a5\u53d7\u957f\u5ea6\u90fd\u662f\u6700\u9ad8\u7684\uff0c\u660e\u663e\u4f18\u4e8e\u5176\u4ed6\u65b9\u6cd5\u3002<\/p>\n

    \u5e94\u7528<\/h4>\n

    EAGLE-3 \u53d1\u5e03\u7b2c\u4e00\u5929\u5c31\u88ab\u96c6\u6210\u5230 SGLang \u4e2d\u3002\u5728\u751f\u4ea7\u7ea7\u6846\u67b6\u4e2d\uff0cEAGLE-3 \u4e5f\u6709\u6570\u500d\u52a0\u901f\u6548\u679c\u3002\u4ee5\u4e0b\u5b9e\u9a8c\u7531 SGLang \u56e2\u961f\u63d0\u4f9b\u5e76\u4ee5 LLaMA 3.1 8B\uff08batch size=1, 1x H100\uff09\u4e3a\u4f8b\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u6295\u673a\u91c7\u6837\u5f80\u5f80\u88ab\u8ba4\u4e3a\u5728\u5927 batch size \u4e0b\u4f1a\u964d\u4f4e\u541e\u5410\u91cf\u3002\u4f46\u662f\u5728 SGLang \u8fd9\u4e00\u751f\u4ea7\u7ea7\u6846\u67b6\u4e0b\uff0cEAGLE-3 \u5728 batch size \u4e3a 64 \u65f6\u4ecd\u53ef\u4ee5\u63d0\u9ad8 38% \u7684\u541e\u5410\u91cf\uff0c\u800c EAGLE \u5728 batch size \u4e3a 24 \u65f6\u5c31\u5bfc\u81f4\u541e\u5410\u91cf\u4e0b\u964d\u3002\u8fd9\u91cc 1.00x \u4ee5 SGLang (w\/o speculative) \u7684\u541e\u5410\u91cf\u4f5c\u4e3a\u57fa\u51c6\u3002\u4ee5\u4e0b\u5b9e\u9a8c\u7531 SGLang \u56e2\u961f\u63d0\u4f9b\u5e76\u4ee5 LLaMA 3.1 8B\uff081x H100\uff09\u4e3a\u4f8b\u3002<\/p>\n

    \"\u56fe\u7247\"<\/p>\n

    \u4f5c\u8005\u4ecb\u7ecd<\/h4>\n

    \u674e\u5809\u6656\uff1a\u5317\u4eac\u5927\u5b66\u667a\u80fd\u5b66\u9662\u7855\u58eb\uff0c\u6ed1\u94c1\u5362\u5927\u5b66\u8bbf\u95ee\u5b66\u8005\uff0c\u53d7\u5f20\u5f18\u626c\u8001\u5e08\u548c\u5f20\u8d85\u8001\u5e08\u6307\u5bfc\uff0c\u7814\u7a76\u65b9\u5411\u4e3a\u5927\u6a21\u578b\u52a0\u901f\u548c\u5bf9\u9f50\u3002<\/p>\n

    \u9b4f\u82b3\u82b8\uff1a\u5fae\u8f6f\u4e9a\u7814\u9662\u7814\u7a76\u5458\uff0c\u7814\u7a76\u65b9\u5411\u4e3a\u5177\u8eab\u667a\u80fd\u3001\u56fe\u50cf\u751f\u6210\u548c AI agents\u3002<\/p>\n

    \u5f20\u8d85\uff1a\u5317\u4eac\u5927\u5b66\u667a\u80fd\u5b66\u9662\u7814\u7a76\u5458\uff0c\u7814\u7a76\u65b9\u5411\u4e3a\u8ba1\u7b97\u673a\u89c6\u89c9\u548c\u5927\u6a21\u578b\u52a0\u901f\u3002<\/p>\n

    \u5f20\u5f18\u626c\uff1a\u6ed1\u94c1\u5362\u5927\u5b66\u8ba1\u7b97\u673a\u5b66\u9662\u3001\u5411\u91cf\u7814\u7a76\u9662\u52a9\u7406\u6559\u6388\uff0c\u7814\u7a76\u65b9\u5411\u4e3a\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f\u3001AI \u5b89\u5168\u548c\u4e16\u754c\u6a21\u578b\u3002<\/p>\n

    \u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n<\/div>\n<\/div>\n

    <\/div>\n

    <\/g><\/svg><\/i> \"Loading\"<\/p>\n

    <\/div>\n","protected":false},"excerpt":{"rendered":"

    \u8fd1\u65e5\uff0cEAGLE \u56e2\u961f\u7684\u65b0\u4f5c\u300aEAGLE-3: Scaling up Inference Accelerati […]<\/p>\n

    <\/div>\n

    <\/g><\/svg><\/i> \"Loading\"<\/p>\n

    <\/div>\n","protected":false},"author":56,"featured_media":25391,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23,17,20,80],"tags":[124,887,1088],"class_list":["post-25389","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23","category-17","category-20","category-80","tag-ai","tag-887","tag-1088"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25389","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=25389"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25389\/revisions"}],"predecessor-version":[{"id":25393,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/25389\/revisions\/25393"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/25391"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=25389"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=25389"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=25389"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}