{"id":23012,"date":"2024-11-11T11:44:20","date_gmt":"2024-11-11T03:44:20","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=23012"},"modified":"2024-11-11T11:44:20","modified_gmt":"2024-11-11T03:44:20","slug":"%e8%85%be%e8%ae%af%e8%af%ad%e9%9f%b3%e5%90%88%e6%88%90%e6%8a%80%e6%9c%af%ef%bc%9a%e6%a8%a1%e5%9e%8b%e4%bc%98%e5%8c%96%e4%b8%8e%e6%8e%a8%e7%90%86%e5%8a%a0%e9%80%9f%e5%ae%9e%e8%b7%b5","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2024\/11\/11\/23012\/","title":{"rendered":"\u817e\u8baf\u8bed\u97f3\u5408\u6210\u6280\u672f\uff1a\u6a21\u578b\u4f18\u5316\u4e0e\u63a8\u7406\u52a0\u901f\u5b9e\u8df5"},"content":{"rendered":"<div class=\"article-desc\">\u968f\u7740\u4eba\u5de5\u667a\u80fd\u6280\u672f\u7684\u4e0d\u65ad\u8fdb\u6b65\uff0c\u8bed\u97f3\u5408\u6210\u6280\u672f\u5728\u6e38\u620f\u548c\u5a31\u4e50\u9886\u57df\u626e\u6f14\u7740\u8d8a\u6765\u8d8a\u91cd\u8981\u7684\u89d2\u8272\u3002\u672c\u6b21\u5206\u4eab\u9898\u76ee\u4e3a\u201c\u817e\u8baf\u6e38\u620f\u77e5\u51e0\u8bed\u97f3\u5408\u6210\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f\u5b9e\u8df5\u201d\uff0c\u4e3b\u8981\u4ecb\u7ecd\u817e\u8baf\u5728\u8bed\u97f3\u5408\u6210\u9886\u57df\u7684\u4ea7\u54c1\u5c55\u793a\u3001\u6a21\u578b\u7ed3\u6784\u5206\u6790\u3001\u63a8\u7406\u52a0\u901f\u65b9\u6848\u4ee5\u53ca\u672a\u6765\u5c55\u671b\u3002<\/div>\n<div id=\"postspictures\" class=\"article-content\">\n<div id=\"container\" class=\"container am-engine\" data-v-1d7a5742=\"\" data-element=\"root\">\n<h2>\u4e00\u3001\u80cc\u666f-\u4ea7\u54c1\u5c55\u793a<\/h2>\n<p>\u9996\u5148\uff0c\u8ba9\u6211\u4eec\u6765\u770b\u4e00\u4e0b\u817e\u8baf\u81ea\u7814\u7684\u77e5\u97f3\u8bed\u97f3\u5927\u6a21\u578b\u5728\u8bed\u97f3\u5408\u6210\u9886\u57df\u7684\u5e94\u7528\u5c55\u793a\u3002\u8be5\u6a21\u578b\u80fd\u591f\u63d0\u4f9b\u66f4\u81ea\u7136\u3001\u97f5\u5f8b\u4e30\u5bcc\u4e14\u5b9e\u65f6\u6027\u66f4\u5f3a\u7684\u8bed\u97f3\u5408\u6210\u4f53\u9a8c\u3002\u5176\u4e24\u5927\u663e\u8457\u4f18\u70b9\u5982\u4e0b\uff1a<\/p>\n<ul data-id=\"ua73dd4b-5HYa7Tcc\">\n<li data-id=\"ld70c578-jCV3IXOk\">\u58f0\u97f3\u590d\u523b\uff1a\u4ec5\u9700 10 \u79d2\u97f3\u9891\u5373\u53ef\u5b8c\u6210\u58f0\u97f3\u590d\u523b\u3002<\/li>\n<li data-id=\"ld70c578-iiV4nLY7\">\u5b9e\u65f6\u6027\u80fd\uff1a\u901a\u8fc7\u52a0\u901f\u4f18\u5316\uff0c\u5176\u5b9e\u65f6\u7387\u7ea6\u4e3a 0.085\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s9.51cto.com\/oss\/202411\/11\/c35c6340591811cd75b587eacb1c354bf3e8ba.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p>\u6211\u4eec\u7684\u4ea7\u54c1\u4e3b\u8981\u5206\u4e3a\u4e09\u79cd\u5f62\u6001\uff1a<\/p>\n<ul data-id=\"ua73dd4b-4V2RHKGQ\">\n<li data-id=\"ld70c578-IIgGU72d\">\u6587\u672c\u95ee\u7b54\uff1a<br \/>\n\u8f93\u5165\u65b9\u5f0f\uff1a\u73a9\u5bb6\u901a\u8fc7\u6587\u672c\u8f93\u5165\u95ee\u9898\u6216\u6307\u4ee4\u3002<br \/>\n\u8f93\u51fa\u65b9\u5f0f\uff1a\u7cfb\u7edf\u5229\u7528\u4eba\u5de5\u667a\u80fd\u6280\u672f\u5904\u7406\u540e\uff0c\u8fd4\u56de\u76f8\u5e94\u7684\u6587\u672c\u56de\u7b54\u3002<\/li>\n<li data-id=\"ld70c578-o73mB39Q\">\u8bed\u97f3\u4ea4\u4e92\uff1a<br \/>\n\u8f93\u5165\u65b9\u5f0f\uff1a\u73a9\u5bb6\u901a\u8fc7\u8bed\u97f3\u4e0e AI \u8bed\u97f3\u52a9\u624b\u8fdb\u884c\u4ea4\u6d41\u3002<br \/>\n\u8f93\u51fa\u65b9\u5f0f\uff1a\u7cfb\u7edf\u7ecf\u8fc7 AI \u5904\u7406\u540e\uff0c\u4ee5\u8bed\u97f3\u5f62\u5f0f\u56de\u5e94\u73a9\u5bb6\u3002<\/li>\n<li data-id=\"ld70c578-CDWG3PJ7\">\u591a\u6a21\u6001 NPC\uff08\u975e\u73a9\u5bb6\u89d2\u8272\uff09\uff1a<br \/>\n\u8f93\u5165\u65b9\u5f0f\uff1a\u73a9\u5bb6\u53ef\u4ee5\u901a\u8fc7\u6587\u672c\u6216\u8bed\u97f3\u4e0e NPC \u8fdb\u884c\u4e92\u52a8\u3002<br \/>\n\u8f93\u51fa\u65b9\u5f0f\uff1a\u7cfb\u7edf\u7ecf\u8fc7 AI \u5904\u7406\u540e\uff0cNPC \u53ef\u4ee5\u8fd4\u56de\u6587\u672c\u3001\u8bed\u97f3\u3001\u80a2\u4f53\u52a8\u4f5c\u3001\u9762\u90e8\u8868\u60c5\u548c\u53e3\u578b\u52a8\u4f5c\u7b49\u591a\u79cd\u5f62\u5f0f\u7684\u6570\u636e\uff0c\u63d0\u4f9b\u66f4\u52a0\u4e30\u5bcc\u548c\u81ea\u7136\u7684\u4ea4\u4e92\u4f53\u9a8c\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e09\u79cd\u4ea7\u54c1\u5f62\u6001\u5206\u522b\u6db5\u76d6\u4e86\u6587\u672c\u3001\u8bed\u97f3\u4ee5\u53ca\u591a\u6a21\u6001\u4ea4\u4e92\uff0c\u65e8\u5728\u4e3a\u7528\u6237\u63d0\u4f9b\u591a\u6837\u5316\u4e14\u9ad8\u5ea6\u6c89\u6d78\u5f0f\u7684\u4e92\u52a8\u4f53\u9a8c\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u4e8c\u3001\u6a21\u578b\u7ed3\u6784\u9009\u578b\u4e0e\u5206\u6790<\/h2>\n<p>\u63a5\u4e0b\u6765\uff0c\u6211\u5c06\u8be6\u7ec6\u8bb2\u89e3\u6211\u4eec\u8bed\u97f3\u5408\u6210\u6a21\u578b\u7684\u7ed3\u6784\u9009\u578b\u548c\u5206\u6790\u3002\u4f20\u7edf\u7684\u8bed\u97f3\u5408\u6210\u65b9\u6848\u901a\u5e38\u5305\u62ec\u4ee5\u4e0b\u6b65\u9aa4\uff1a<\/p>\n<ul data-id=\"ua73dd4b-GJM8NAdj\">\n<li data-id=\"ld70c578-76A7fIUU\">\u8f93\u5165\u6587\u672c\u3002<\/li>\n<li data-id=\"ld70c578-mFUUK68s\">\u901a\u8fc7\u58f0\u5b66\u6a21\u578b\u751f\u6210\u6885\u5c14\u8c31\u56fe\uff08Mel-spectrogram\uff09\u3002<\/li>\n<li data-id=\"ld70c578-gR9sahZV\">\u901a\u8fc7\u58f0\u7801\u5668\uff08Vocoder\uff09\u5c06\u6885\u5c14\u8c31\u56fe\u8f6c\u6362\u4e3a\u8bed\u97f3\u3002<\/li>\n<\/ul>\n<p>\u7136\u800c\uff0c\u8fd9\u79cd\u4f20\u7edf\u65b9\u6848\u5b58\u5728\u4e00\u4e2a\u663e\u8457\u7f3a\u70b9\uff1a\u9700\u8981\u5927\u91cf\u7684\u8bad\u7ec3\u8bed\u6599\u3002\u5728\u6e38\u620f\u573a\u666f\u4e2d\uff0c\u4f8b\u5982 NPC\uff08\u975e\u73a9\u5bb6\u89d2\u8272\uff09\uff0c\u5f80\u5f80\u96be\u4ee5\u83b7\u53d6\u5982\u6b64\u5927\u91cf\u7684\u8bad\u7ec3\u6570\u636e\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u7684\u7814\u7a76\u65b9\u5411\u8f6c\u5411\u4e86\u57fa\u4e8e\u8bed\u8a00\u6a21\u578b\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n<h3>1.\u00a0\u8bed\u97f3\u5408\u6210\u5927\u6a21\u578b\u7ed3\u6784<\/h3>\n<p>\u65b0\u65b9\u6848\u7684\u4e3b\u8981\u6d41\u7a0b\u5982\u4e0b\uff1a<\/p>\n<ul data-id=\"ua73dd4b-57PDZEUT\">\n<li data-id=\"ld70c578-LF5eE78a\">\u8f93\u5165\u6587\u672c\uff1a<br \/>\n\u8f93\u5165\u4e00\u6bb5\u6587\u672c\u3002<\/li>\n<li data-id=\"ld70c578-oUi1eTlQ\">Tokenization\uff1a<br \/>\n\u5bf9\u8f93\u5165\u6587\u672c\u548c\u9700\u8981\u514b\u9686\u7684\u8bed\u97f3\u5206\u522b\u8fdb\u884c Token \u5316\u5904\u7406\uff0c\u751f\u6210\u6587\u672c Token \u548c\u8bed\u97f3 Token\u3002<\/li>\n<li data-id=\"ld70c578-8f2pKa2k\">\u8fde\u63a5\u4e0e\u7f16\u7801\uff1a<br \/>\n\u5c06\u6587\u672c Token \u548c\u8bed\u97f3 Token \u8fde\u63a5\u5728\u4e00\u8d77\u3002<br \/>\n\u901a\u8fc7\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801\uff08Rotary Position Encoding\uff09\u5bf9\u6587\u672c\u548c\u97f3\u9891\u5206\u522b\u8fdb\u884c\u7f16\u7801\uff0c\u751f\u6210\u6700\u7ec8\u7684\u5d4c\u5165\u5411\u91cf\uff08Embedding\uff09\u3002<\/li>\n<li data-id=\"ld70c578-F2XZrQZP\">\u81ea\u56de\u5f52\u6a21\u578b\uff08AR Model\uff09\uff1a<br \/>\n\u5c06\u5d4c\u5165\u5411\u91cf\u8f93\u5165\u5230 AR \u6a21\u578b\u4e2d\uff0c\u9010\u6b65\u751f\u6210\u7b2c\u4e00\u5c42\u97f3\u7d20 Token\u3002<br \/>\nAR\u00a0\u6a21\u578b\u7684\u7ed3\u6784\u4e0e\u4f20\u7edf\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5927\u6a21\u578b\uff08\u5982 GPT \u7c7b\u6a21\u578b\uff09\u7c7b\u4f3c\uff0c\u91c7\u7528 Transformer \u67b6\u6784\u3002<\/li>\n<li data-id=\"ld70c578-uTDMRIdU\">\u975e\u81ea\u56de\u5f52\u6a21\u578b\uff08NAR Model\uff09\uff1a<br \/>\n\u5c06 AR \u6a21\u578b\u751f\u6210\u7684\u97f3\u7d20 Token \u8f93\u5165\u5230 NAR \u6a21\u578b\u4e2d\uff0c\u751f\u6210\u6240\u6709\u8bed\u97f3 Token\u3002<br \/>\nCodec Decoder\uff1a<br \/>\n\u6700\u540e\uff0c\u901a\u8fc7 Codec Decoder \u5c06\u8bed\u97f3 Token \u8f6c\u6362\u4e3a\u6700\u7ec8\u7684\u8bed\u97f3\u8f93\u51fa\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s9.51cto.com\/oss\/202411\/11\/310cdac41baf7049fbe107c37434a9749fdedb.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<h3>2.\u00a0\u9762\u4e34\u7684\u6311\u6218<\/h3>\n<p>\u9ad8\u5e76\u53d1\u573a\u666f\u7684\u6311\u6218\uff1a\u5728\u9ad8\u5e76\u53d1\u73af\u5883\u4e0b\uff0c\u6a21\u578b\u9700\u8981\u9ad8\u6548\u5904\u7406\u5927\u91cf\u8bf7\u6c42\u3002<\/p>\n<p>\u5b9e\u65f6\u7387\u8981\u6c42\uff1a\u7279\u522b\u662f\u5728\u6e38\u620f\u573a\u666f\u4e2d\uff0c\u9700\u8981\u5b9e\u65f6\u5904\u7406\u8bed\u97f3\u5e76\u89e6\u8fbe\u73a9\u5bb6\uff0c\u5bf9\u5b9e\u65f6\u7387\u7684\u8981\u6c42\u975e\u5e38\u9ad8\u3002<\/p>\n<p>\u9488\u5bf9\u4e0a\u8ff0\u6311\u6218\uff0c\u7b2c\u4e09\u90e8\u5206\u6211\u4eec\u5c06\u8be6\u7ec6\u4ecb\u7ecd\u6a21\u578b\u63a8\u7406\u7684\u52a0\u901f\u65b9\u6848\uff0c\u4ee5\u63d0\u9ad8\u7cfb\u7edf\u7684\u5e76\u53d1\u5904\u7406\u80fd\u529b\u548c\u5b9e\u65f6\u6027\u80fd\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u4e09\u3001\u6a21\u578b\u63a8\u7406\u52a0\u901f\u65b9\u6848<\/h2>\n<p>\u5728\u9009\u62e9\u52a0\u901f\u65b9\u6848\u65f6\uff0c\u6211\u4eec\u9996\u5148\u8003\u8651\u501f\u9274\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u9886\u57df\u7684\u4e00\u4e9b\u6210\u719f\u52a0\u901f\u65b9\u6cd5\uff0c\u5e76\u5c06\u5176\u5e94\u7528\u5230\u8bed\u97f3\u5408\u6210\u5927\u6a21\u578b\u4e0a\u3002\u5728 NLP \u9886\u57df\u7684\u5927\u6a21\u578b\u4e2d\uff0c\u5df2\u7ecf\u6709\u8bb8\u591a\u6210\u719f\u7684\u63a8\u7406\u52a0\u901f\u65b9\u6cd5\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u53ef\u4ee5\u5206\u4e3a\u5de5\u7a0b\u4e0a\u7684\u4f18\u5316\uff08KV Cache\uff0cprefix KV Cache\uff0cflash attention\uff0cflash decode\u548cpage attention\uff09\u548c\u504f\u7b97\u6cd5\u4e0a\u7684\u4f18\u5316\uff08\u6295\u673a\u91c7\u6837\uff0cint4\/int8 \u91cf\u5316\u7b49\uff09\u3002\u8fd9\u7ed9\u6211\u4eec\u5e26\u6765\u4e86\u975e\u5e38\u591a\u7684\u542f\u53d1\u3002<\/p>\n<h3>1.\u00a0KV Cache<\/h3>\n<p>\u4e3a\u4e86\u5c06 KV Cache \u6280\u672f\u5e94\u7528\u4e8e\u8bed\u97f3\u5408\u6210\u5927\u6a21\u578b\uff0c\u6211\u4eec\u9700\u8981\u786e\u4fdd Attention Mask \u7684\u8bbe\u8ba1\u6ee1\u8db3\u7279\u5b9a\u6761\u4ef6\u3002\u8fd9\u4e9b\u6761\u4ef6\u4f7f\u5f97 KV Cache \u80fd\u591f\u6709\u6548\u5730\u7f13\u5b58\u4e4b\u524d\u7684\u8ba1\u7b97\u7ed3\u679c\uff0c\u4ece\u800c\u51cf\u5c11\u5197\u4f59\u8ba1\u7b97\u5e76\u63d0\u9ad8\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n<h4>\uff081\uff09\u4f7f\u7528 KV Cache \u7684\u6761\u4ef6<\/h4>\n<ul data-id=\"ua73dd4b-j3EDdJAP\">\n<li data-id=\"ld70c578-o73O2fGJ\">\u00a0Attention\u00a0\u8f93\u51fa\u7684\u7b2c n \u884c\u53ea\u4e0e\u7b2c n \u4e2a Query (Q)\u76f8\u5173\uff1a\u8fd9\u610f\u5473\u7740\u6bcf\u4e2a Token \u7684\u8f93\u51fa\u53ea\u4f9d\u8d56\u4e8e\u5f53\u524d Token \u7684 Query \u548c\u4e4b\u524d\u6240\u6709 Token \u7684 Key (K) \u548c Value (V)\u3002<\/li>\n<li data-id=\"ld70c578-U0uZFsLz\">\u7b2c 1 \u5230\u7b2c n \u4e2a Token \u7684 Attention \u8ba1\u7b97\u5305\u542b\u7b2c 1 \u5230\u7b2c n-1 \u4e2a Token \u7684Attention \u8ba1\u7b97\uff1a\u8fd9\u610f\u5473\u7740\u6bcf\u6b21\u8ba1\u7b97\u65b0\u7684 Token \u65f6\uff0c\u53ef\u4ee5\u590d\u7528\u4e4b\u524d Token \u7684 K \u548c V \u503c\uff0c\u800c\u4e0d\u9700\u8981\u91cd\u65b0\u8ba1\u7b97\u3002<\/li>\n<li data-id=\"ld70c578-Mykws7gV\">\u6bcf\u6b21 Attention \u7684\u8ba1\u7b97\u90fd\u4f7f\u7528\u524d\u9762\u7684 K \u548c V\uff1a\u6bcf\u6b21\u751f\u6210\u65b0\u7684 Token \u65f6\uff0c\u53ea\u9700\u8ba1\u7b97\u65b0 Token \u7684 K \u548c V \u503c\uff0c\u5e76\u7ed3\u5408\u4e4b\u524d\u7f13\u5b58\u7684 K \u548c V \u503c\u8fdb\u884c\u8ba1\u7b97\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s2.51cto.com\/oss\/202411\/11\/553271b438e305a15dd5220fdea22e696c4328.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<h4>\uff082\uff09\u8bed\u97f3 AR \u6a21\u578b\u4e2d\u7684 Attention Mask \u8bbe\u8ba1<\/h4>\n<p>\u5728\u8bed\u97f3\u81ea\u56de\u5f52\uff08AR\uff09\u6a21\u578b\u4e2d\uff0cAttention Mask \u7684\u8bbe\u8ba1\u901a\u5e38\u6ee1\u8db3\u4e0a\u8ff0\u6761\u4ef6\uff0c\u56e0\u6b64\u53ef\u4ee5\u6709\u6548\u5730\u4f7f\u7528 KV Cache\u3002\u6ce8\u610f\u529b\u63a9\u7801\uff08attention_mask\uff09\u548c\u63a9\u7801\u540e\u7684 QK \u77e9\u9635\uff08Mask(QK)\uff09\u7684\u53d8\u5316\u8fc7\u7a0b\uff1a\u5de6\u4fa7\u5c55\u793a\u4e86\u4e00\u4e2a\u6ce8\u610f\u529b\u63a9\u7801\u7684\u4f8b\u5b50\uff0c\u5176\u4e2d\u6bcf\u4e2a\u5143\u7d20\u8868\u793a\u5bf9\u5e94\u4f4d\u7f6e\u7684 token \u662f\u5426\u53c2\u4e0e\u6ce8\u610f\u529b\u8ba1\u7b97\u3002\u4e2d\u95f4\u5c55\u793a\u4e86\u63a9\u7801\u540e\u7684 Q*K \u77e9\u9635\uff0c\u5176\u4e2d\u975e\u96f6\u5143\u7d20\u8868\u793a\u6709\u6548\u7684\u6ce8\u610f\u529b\u6743\u91cd\u3002\u5f53\u751f\u6210\u7b2c n \u4e2a token \u65f6\uff0c\u6ce8\u610f\u529b\u63a9\u7801\u4f1a\u53d1\u751f\u76f8\u5e94\u7684\u53d8\u5316\uff0c\u65b0\u7684 token \u4f1a\u88ab\u52a0\u5165\u5230\u63a9\u7801\u4e2d\u3002\u53f3\u4fa7\u5c55\u793a\u4e86\u66f4\u65b0\u540e\u7684\u6ce8\u610f\u529b\u63a9\u7801\u548c\u76f8\u5e94\u7684\u63a9\u7801\u540e\u7684 Q*K \u77e9\u9635\u3002<\/p>\n<p>\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0c\u6bcf\u6b21\u751f\u6210\u65b0 token \u65f6\uff0c\u53ea\u9700\u8981\u8ba1\u7b97\u65b0\u589e\u7684\u90e8\u5206\uff0c\u800c\u4e0d\u9700\u8981\u91cd\u65b0\u8ba1\u7b97\u6574\u4e2a\u5e8f\u5217\u7684\u6ce8\u610f\u529b\u6743\u91cd\uff0c\u4ece\u800c\u5927\u5927\u51cf\u5c11\u4e86\u8ba1\u7b97\u91cf\uff0c\u63d0\u9ad8\u4e86\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s5.51cto.com\/oss\/202411\/11\/b1919ef5543037997a28522c0f4b3beb51c6c4.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<h3>2.\u00a0GQA<\/h3>\n<p>\u5728\u9884\u586b\u5145\uff08Prefill\uff09\u9636\u6bb5\uff0c\u8ba1\u7b97\u74f6\u9888\u662f\u4e3b\u8981\u7684\u9650\u5236\u56e0\u7d20\u3002\u8fd9\u4e00\u9636\u6bb5\u901a\u5e38\u6d89\u53ca\u5927\u91cf\u7684\u5e76\u884c\u8ba1\u7b97\uff0c\u7279\u522b\u662f\u5728\u5904\u7406\u521d\u59cb\u8f93\u5165\u65f6\uff0c\u6a21\u578b\u9700\u8981\u5bf9\u6574\u4e2a\u5e8f\u5217\u8fdb\u884c\u7f16\u7801\uff0c\u751f\u6210\u6240\u6709 Token \u7684 Key (K) \u548c Value (V) \u503c\u3002\u8fd9\u79cd\u5927\u89c4\u6a21\u7684\u8ba1\u7b97\u9700\u6c42\u4f7f\u5f97\u8ba1\u7b97\u8d44\u6e90\u6210\u4e3a\u4e3b\u8981\u74f6\u9888\u3002<\/p>\n<p>\u800c\u5728\u89e3\u7801\uff08Decode\uff09\u9636\u6bb5\uff0c\u5185\u5b58\u8bbf\u95ee\u74f6\u9888\u5219\u6210\u4e3a\u4e3b\u5bfc\u3002\u5728\u89e3\u7801\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u9010\u6b65\u751f\u6210\u65b0\u7684 Token\uff0c\u5e76\u4e14\u6bcf\u6b21\u751f\u6210\u65b0 Token \u65f6\u90fd\u9700\u8981\u8bbf\u95ee\u4e4b\u524d\u751f\u6210\u7684\u6240\u6709 Token \u7684 K \u548c V \u503c\u3002\u9891\u7e41\u7684\u5185\u5b58\u8bbf\u95ee\u5bfc\u81f4\u4e86\u8f83\u9ad8\u7684\u5ef6\u8fdf\uff0c\u4f7f\u5f97\u5185\u5b58\u5e26\u5bbd\u6210\u4e3a\u4e3b\u8981\u7684\u6027\u80fd\u74f6\u9888\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s7.51cto.com\/oss\/202411\/11\/c5d407634b3baccd4a2484735e8a59ab4c98c0.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p>\u76f8\u6bd4\u4e8e\u4f7f\u7528 KV Cache\u3001INT8\/FP8 \u91cf\u5316\u7b49\u65b9\u6cd5\uff0c\u9009\u62e9\u7ec4\u91cf\u5316\u6ce8\u610f\u529b\uff08Grouped Query Attention, GQA\uff09\u53ef\u4ee5\u63d0\u4f9b\u66f4\u53ef\u63a7\u7684\u538b\u7f29\u7387\u3002GQA \u901a\u8fc7\u5c06\u591a\u4e2a\u67e5\u8be2\u5934\uff08Query Heads\uff09\u7ec4\u5408\u5728\u4e00\u8d77\uff0c\u51cf\u5c11\u4e86\u6a21\u578b\u4e2d\u6ce8\u610f\u529b\u5934\u7684\u6570\u91cf\uff0c\u4ece\u800c\u964d\u4f4e\u4e86\u8ba1\u7b97\u590d\u6742\u5ea6\u548c\u5185\u5b58\u8bbf\u95ee\u91cf\u3002\u8fd9\u79cd\u65b9\u6cd5\u53ef\u4ee5\u5728\u4fdd\u8bc1\u6a21\u578b\u6548\u679c\u7684\u540c\u65f6\uff0c\u663e\u8457\u63d0\u9ad8\u63a8\u7406\u6548\u7387\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u901a\u8fc7\u5c06\u6ce8\u610f\u529b\u5934\u6570\u91cf\u4ece 16 \u51cf\u5c11\u5230 4\uff0c\u6211\u4eec\u4e0d\u4ec5\u51cf\u5c11\u4e86\u6a21\u578b\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u8fd8\u964d\u4f4e\u4e86\u5185\u5b58\u8bbf\u95ee\u7684\u9700\u6c42\u3002\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u8fd9\u79cd\u4f18\u5316\u53ef\u4ee5\u4f7f\u63a8\u7406\u8017\u65f6\u964d\u4f4e\u00a0 20%\u3002\u8fd9\u8868\u660e\uff0cGQA \u662f\u4e00\u79cd\u6709\u6548\u7684\u624b\u6bb5\uff0c\u80fd\u591f\u5728\u4fdd\u6301\u6a21\u578b\u6027\u80fd\u7684\u540c\u65f6\uff0c\u663e\u8457\u63d0\u5347\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s3.51cto.com\/oss\/202411\/11\/15043f89817e8d353bf126ba5472503f4d9779.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<h4>3.\u00a0BPE<\/h4>\n<p>\u5728\u5f15\u5165 KV Cache \u673a\u5236\u540e\uff0cAR \u6a21\u578b\u88ab\u5212\u5206\u4e3a\u4e24\u4e2a\u4e3b\u8981\u9636\u6bb5\uff1a\u9884\u586b\u5145\uff08Prefill\uff09\u9636\u6bb5\u548c\u89e3\u7801\uff08Decode\uff09\u9636\u6bb5\u3002\u9884\u586b\u5145\u9636\u6bb5\u8d1f\u8d23\u521d\u59cb\u5316\u6a21\u578b\u7684\u72b6\u6001\uff0c\u5e76\u4e3a\u540e\u7eed\u7684\u89e3\u7801\u8fc7\u7a0b\u51c6\u5907\u5fc5\u8981\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\uff1b\u800c\u89e3\u7801\u9636\u6bb5\u5219\u57fa\u4e8e\u8fd9\u4e9b\u4e0a\u4e0b\u6587\u4fe1\u606f\u9010\u4e2a\u751f\u6210\u97f3\u9891\u6837\u672c\u6216 token\u3002\u5bf9\u4e8e\u751f\u6210\u4e00\u6bb5\u65f6\u957f\u4e3a 10 \u79d2\u7684\u9ad8\u8d28\u91cf\u97f3\u9891\u800c\u8a00\uff0c\u91c7\u7528\u4f20\u7edf\u65b9\u6cd5\u901a\u5e38\u9700\u8981 AR \u6a21\u578b\u4ea7\u751f\u5927\u7ea6 500 \u4e2a token\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s8.51cto.com\/oss\/202411\/11\/9558fbc60b1a279d3d7770d12285b0d886ffc1.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p>NLP\u00a0\u9886\u57df\u4e2d\u5e7f\u6cdb\u91c7\u7528\u5b57\u8282\u5bf9\u7f16\u7801\uff08Byte Pair Encoding, BPE\uff09\u6280\u672f\u6765\u89e3\u51b3 OOV \u95ee\u9898\u3002BPE \u7b97\u6cd5\u9996\u5148\u5c06\u8f93\u5165\u6587\u672c\u4e2d\u7684\u6bcf\u4e2a\u5355\u8bcd\u62c6\u5206\u6210\u5355\u4e2a\u5b57\u7b26\u5e8f\u5217\uff0c\u968f\u540e\u901a\u8fc7\u7edf\u8ba1\u5206\u6790\u627e\u51fa\u51fa\u73b0\u9891\u7387\u6700\u9ad8\u7684\u5b57\u7b26\u5bf9\uff0c\u5e76\u7528\u4e00\u4e2a\u65b0\u7684\u590d\u5408\u7b26\u53f7\u4ee3\u66ff\u8fd9\u5bf9\u5b57\u7b26\u3002\u8fd9\u4e00\u66ff\u6362\u6b65\u9aa4\u4f1a\u91cd\u590d\u8fdb\u884c\uff0c\u76f4\u81f3\u8fbe\u5230\u9884\u8bbe\u7684\u6700\u5927\u5faa\u73af\u6b21\u6570\u6216\u6ee1\u8db3\u5176\u4ed6\u505c\u6b62\u6761\u4ef6\u4e3a\u6b62\u3002\u8fd9\u79cd\u65b9\u6cd5\u6709\u6548\u5730\u589e\u52a0\u4e86\u8bcd\u5178\u5bb9\u91cf\u7684\u540c\u65f6\uff0c\u4fdd\u6301\u4e86\u76f8\u5bf9\u8f83\u5c0f\u7684\u8bcd\u6c47\u8868\u5927\u5c0f\uff0c\u6709\u52a9\u4e8e\u63d0\u9ad8\u6a21\u578b\u8bad\u7ec3\u6548\u7387\u53ca\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s3.51cto.com\/oss\/202411\/11\/36d36b63334448e2663629845bbbe917211334.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p>\u4e3a\u4e86\u76f4\u63a5\u4f7f\u7528 NLP BPE\uff0c\u5148\u5c06\u539f\u59cb\u97f3\u9891\u6570\u636e\u8f6c\u6362\u6210\u4e00\u7cfb\u5217\u79bb\u6563\u5316\u7684 audio token\uff0c\u7136\u540e\u5c06\u8fd9\u4e9b token \u4e0e\u552f\u4e00\u7684 Unicode \u5b57\u7b26\u5efa\u7acb\u4e00\u4e00\u5bf9\u5e94\u5173\u7cfb\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u6bcf\u4e00\u4e2a\u7279\u5b9a\u7684 Unicode \u5b57\u7b26\u5b9e\u9645\u4e0a\u5c31\u7f16\u7801\u4e86\u4e00\u4e2a\u6216\u4e00\u7ec4\u76f8\u5173\u7684\u97f3\u9891\u4fe1\u53f7\u7247\u6bb5\u3002\u8fd9\u79cd\u505a\u6cd5\u4e0d\u4ec5\u7b80\u5316\u4e86\u6a21\u578b\u7ed3\u6784\uff0c\u8fd8\u6781\u5927\u5730\u63d0\u5347\u4e86\u5904\u7406\u6548\u7387\u3002\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u901a\u8fc7\u5e94\u7528\u4e0a\u8ff0\u4f18\u5316\u7b56\u7565\uff0c\u5408\u6210\u76f8\u540c\u957f\u5ea6\u5373 10 \u79d2\u949f\u97f3\u9891\u6240\u9700\u7684 token \u6570\u91cf\u53ef\u7531\u539f\u6765\u7684\u7ea6 500 \u4e2a\u51cf\u5c11\u81f3\u7ea6 170 \u4e2a\u5de6\u53f3\uff0c\u663e\u8457\u964d\u4f4e\u4e86\u8ba1\u7b97\u590d\u6742\u5ea6\u5e76\u52a0\u901f\u4e86\u6574\u4e2a\u751f\u6210\u6d41\u7a0b\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s4.51cto.com\/oss\/202411\/11\/7109fb809bdcf77a404653cff709a79ad82243.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<h3>4.\u00a0\u8fde\u7eed\u6027\u6279\u5904\u7406<\/h3>\n<p>LLM\u00a0\u4e2d\u4f18\u5316\u541e\u5410\u4e3b\u8981\u6709\u4e24\u79cd\u65b9\u5f0f\uff1a<\/p>\n<p>\u6734\u7d20\u6279\u5904\u7406\uff1a\u6bcf\u4e2a\u65f6\u95f4\u6b65\uff08T1 \u5230 T8\uff09\u5206\u522b\u72ec\u7acb\u5730\u5904\u7406\u4e0d\u540c\u7684\u53e5\u5b50\uff08S1 \u5230 S4\uff09\uff0c\u5e76\u4e14\u6bcf\u4e2a\u53e5\u5b50\u5728\u8fbe\u5230\u7ed3\u675f\u6807\u8bb0\uff08END\uff09\u4e4b\u524d\u90fd\u4f1a\u88ab\u5355\u72ec\u5904\u7406\u3002<\/p>\n<p>\u8fde\u7eed\u6027\u6279\u5904\u7406\uff1a\u8bd5\u56fe\u5728\u540c\u4e00\u65f6\u95f4\u6b65\u5185\u540c\u65f6\u5904\u7406\u591a\u4e2a\u53e5\u5b50\uff0c\u4ee5\u5b9e\u73b0\u66f4\u9ad8\u7684\u5e76\u884c\u6027\u548c\u6548\u7387\u3002\u4f8b\u5982\uff0c\u5728\u7b2c\u4e00\u4e2a\u65f6\u95f4\u6b65\uff08T1\uff09\uff0c\u56db\u4e2a\u53e5\u5b50\uff08S1 \u5230 S4\uff09\u90fd\u88ab\u5904\u7406\uff1b\u968f\u540e\u7684\u65f6\u95f4\u6b65\u7ee7\u7eed\u5904\u7406\u8fd9\u4e9b\u53e5\u5b50\u76f4\u5230\u5b83\u4eec\u5404\u81ea\u5b8c\u6210\u3002<\/p>\n<p>\u8fd9\u91cc\u53ef\u4ee5\u76f4\u89c2\u7684\u770b\u5230\u4e24\u79cd\u6279\u5904\u7406\u65b9\u6cd5\u7684\u533a\u522b\uff1a<\/p>\n<p>\u5de6\u4e0a\u89d2\u56fe\u8868\uff1a\u663e\u793a\u4e86\u6734\u7d20\u6279\u5904\u7406\u7684\u7ed3\u679c\uff0c\u6bcf\u4e2a\u65f6\u95f4\u6b65\u4ec5\u5904\u7406\u4e00\u4e2a\u53e5\u5b50\u3002<\/p>\n<p>\u53f3\u4e0a\u89d2\u56fe\u8868\uff1a\u663e\u793a\u4e86\u6734\u7d20\u6279\u5904\u7406\u7684\u6700\u7ec8\u72b6\u6001\uff0c\u6bcf\u4e2a\u53e5\u5b50\u90fd\u5728\u5404\u81ea\u7684\u7ed3\u675f\u70b9\u505c\u6b62\u3002<\/p>\n<p>\u5de6\u4e0b\u89d2\u56fe\u8868\uff1a\u663e\u793a\u4e86\u8fde\u7eed\u6027\u6279\u5904\u7406\u7684\u5f00\u59cb\u9636\u6bb5\uff0c\u6240\u6709\u53e5\u5b50\u5728\u7b2c\u4e00\u4e2a\u65f6\u95f4\u6b65\u90fd\u88ab\u5904\u7406\u3002<\/p>\n<p>\u53f3\u4e0b\u89d2\u56fe\u8868\uff1a\u663e\u793a\u4e86\u8fde\u7eed\u6027\u6279\u5904\u7406\u7684\u6700\u7ec8\u72b6\u6001\uff0c\u4e0d\u540c\u53e5\u5b50\u5728\u4e0d\u540c\u7684\u65f6\u95f4\u6b65\u5b8c\u6210\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s4.51cto.com\/oss\/202411\/11\/161a16582b999550f6405970c703dbab75aced.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p>\u5728\u5927\u578b\u8bed\u97f3\u5408\u6210\u6a21\u578b\u4e2d\uff0c\u53ef\u4ee5\u91c7\u7528\u4e24\u79cd\u4e0d\u540c\u7684\u6279\u5904\u7406\uff08batching\uff09\u7b56\u7565\u3002\u4e0b\u9762\u5206\u522b\u4ecb\u7ecd\u4e86\u8fd9\u4e24\u79cd\u65b9\u6848\u53ca\u5176\u4f18\u7f3a\u70b9\uff1a<\/p>\n<ul data-id=\"ua73dd4b-o1cqEVle\">\n<li data-id=\"ld70c578-EKoSVMCA\">\u8bad\u7ec3\u65f6\u91c7\u7528\u7684\u65b9\u6848\u4e00<\/li>\n<\/ul>\n<p>\u4f18\u70b9\uff1a\u5f53\u9700\u8981\u5bf9\u6587\u672c\u548c\u97f3\u9891\u7279\u5f81\u8fdb\u884c\u7c7b\u4f3c\u5d4c\u5165\uff08embedding, emb\uff09\u7684\u64cd\u4f5c\u65f6\uff0c\u8fd9\u79cd\u65b9\u6848\u63d0\u4f9b\u4e86\u66f4\u52a0\u7b80\u6d01\u76f4\u63a5\u7684\u65b9\u6cd5\u3002\u5b83\u5141\u8bb8\u6a21\u578b\u5206\u522b\u5904\u7406\u6587\u672c\u548c\u97f3\u9891\u6570\u636e\uff0c\u4ece\u800c\u7b80\u5316\u4e86\u7279\u5f81\u5904\u7406\u6d41\u7a0b\u3002<\/p>\n<p>\u7f3a\u70b9\uff1a\u5728\u63a8\u7406\u9636\u6bb5\uff0c\u5c24\u5176\u662f\u5728\u89e3\u7801\u8fc7\u7a0b\u4e2d\u6267\u884c\u6ce8\u610f\u529b\u673a\u5236\u8ba1\u7b97\u65f6\uff0c\u5fc5\u987b\u4e3a\u6bcf\u6b21\u8ba1\u7b97\u63d0\u4f9b\u586b\u5145\u6ce8\u610f\u529b\u63a9\u7801\uff08padding attention mask\uff09\u3002\u8fd9\u589e\u52a0\u4e86\u63a8\u7406\u6846\u67b6\u5b9e\u73b0\u4e0a\u7684\u590d\u6742\u6027\uff0c\u56e0\u4e3a\u9700\u8981\u989d\u5916\u5904\u7406\u8fd9\u4e9b\u63a9\u7801\u4ee5\u786e\u4fdd\u6b63\u786e\u5730\u5ffd\u7565\u6389\u586b\u5145\u90e8\u5206\u3002<\/p>\n<ul data-id=\"ua73dd4b-HfqD20H2\">\n<li data-id=\"ld70c578-oHNqWjop\">\u63a8\u7406\u65f6\u91c7\u7528\u7684\u65b9\u6848\u4e8c<\/li>\n<\/ul>\n<p>\u4f18\u70b9\uff1a\u8be5\u65b9\u6848\u7b80\u5316\u4e86\u6ce8\u610f\u529b\u673a\u5236\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u5e94\u7528\uff0c\u65e0\u9700\u4e3a\u6bcf\u6b21\u8ba1\u7b97\u5355\u72ec\u5b9a\u4e49\u6216\u4f20\u9012\u586b\u5145\u6ce8\u610f\u529b\u63a9\u7801\u3002\u56e0\u6b64\uff0c\u5b83\u53ef\u4ee5\u65e0\u7f1d\u96c6\u6210\u5230\u73b0\u6709\u7684\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u63a8\u7406\u6846\u67b6\u4e2d\uff0c\u63d0\u9ad8\u4e86\u6613\u7528\u6027\u548c\u6548\u7387\u3002<\/p>\n<p>\u7f3a\u70b9\uff1a\u7136\u800c\uff0c\u5728\u9762\u5bf9\u9700\u8981\u540c\u65f6\u5904\u7406\u6587\u672c\u548c\u97f3\u9891\u7279\u5f81\u7684\u60c5\u51b5\u65f6\uff0c\u5982\u8fdb\u884c\u5d4c\u5165\u64cd\u4f5c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u53ef\u80fd\u5bfc\u81f4\u5904\u7406\u6d41\u7a0b\u53d8\u5f97\u66f4\u4e3a\u590d\u6742\u3002\u7531\u4e8e\u4e0d\u80fd\u50cf\u65b9\u6848\u4e00\u90a3\u6837\u76f4\u63a5\u5206\u79bb\u5904\u7406\uff0c\u56e0\u6b64\u53ef\u80fd\u9700\u8981\u8bbe\u8ba1\u989d\u5916\u7684\u903b\u8f91\u6765\u534f\u8c03\u4e0d\u540c\u6a21\u6001\u4e4b\u95f4\u7684\u4ea4\u4e92\uff0c\u589e\u52a0\u4e86\u7cfb\u7edf\u7684\u6574\u4f53\u590d\u6742\u5ea6\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s2.51cto.com\/oss\/202411\/11\/8182b7187bfd975ca633631a3d843a9e69ec6e.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p>\u8fd9\u91cc\u6211\u4eec\u4ecb\u7ecd\u4e00\u4e0b\u7ed3\u5408\u817e\u8baf Trpc \u5fae\u670d\u52a1\u6846\u67b6\uff0c\u5728\u8bed\u97f3\u5408\u6210\u5927\u6a21\u578b\u4e2d\u5b9e\u8df5 continuous batching \u63a8\u7406\u7684\u8fc7\u7a0b\u3002\u8be5\u6846\u67b6\u65e8\u5728\u63d0\u5347\u6a21\u578b\u63a8\u7406\u6027\u80fd\uff0c\u964d\u4f4e\u5ef6\u8fdf\u5e76\u63d0\u9ad8\u8d44\u6e90\u5229\u7528\u7387\u3002<\/p>\n<p>\u5de6\u4fa7\u7684\u67b6\u6784\u56fe\u4ecb\u7ecd\u7684\u662f\u4e00\u4e2a\u5178\u578b\u7684\u5206\u5e03\u5f0f\u7cfb\u7edf\u7ed3\u6784\u3002\u5176\u4e2d\uff0c\u201cmaster\u201d\u8fdb\u7a0b\u4f5c\u4e3a\u4e3b\u63a7\u8282\u70b9\uff0c\u8d1f\u8d23\u4efb\u52a1\u8c03\u5ea6\u548c\u7ba1\u7406\u3002\u4ece\u201cmaster\u201d\u5206\u652f\u51fa\u6765\u7684\u201cDispatcher\u201d\u8fdb\u7a0b\u7528\u4e8e\u63a5\u6536\u7528\u6237\u8bf7\u6c42\u5e76\u5c06\u5b83\u4eec\u5206\u914d\u7ed9\u5404\u4e2a\u5de5\u4f5c\u8fdb\u7a0b\u7ec4\u3002\u5de5\u4f5c\u8fdb\u7a0b\u7ec4\u5305\u542b\u591a\u4e2a\u201cWorker\u201d\u8fdb\u7a0b\uff0c\u5b83\u4eec\u8fd0\u884c\u5728 GPU \u6216 CPU \u4e0a\uff0c\u8d1f\u8d23\u5b9e\u9645\u7684\u6a21\u578b\u63a8\u7406\u4efb\u52a1\u3002\u6b64\u5916\uff0c\u8fd8\u6709\u4e00\u4e2a\u201cTask\u201d\u8fdb\u7a0b\uff0c\u5176\u5185\u90e8\u6709\u4e00\u4e2a\u7f13\u51b2\u533a\uff08buffer\uff09\uff0c\u7528\u4e8e\u5b58\u50a8\u5f85\u5904\u7406\u7684\u4efb\u52a1\u3002\u8fd9\u4e9b\u4efb\u52a1\u7ecf\u8fc7\u524d\u5904\u7406\u540e\uff0c\u8fdb\u5165\u6a21\u578b\u63a8\u7406\u9636\u6bb5\uff0c\u6700\u540e\u8fd4\u56de\u7ed3\u679c\u5e76\u8fdb\u884c\u540e\u5904\u7406\u3002<\/p>\n<p>\u53f3\u4fa7\u7684\u6d41\u7a0b\u56fe\u8be6\u7ec6\u63cf\u8ff0\u4e86\u5177\u4f53\u7684\u6301\u7eed\u6279\u91cf\u63a8\u7406\u8fc7\u7a0b\u3002\u9996\u5148\uff0c\u901a\u8fc7 rpc \u6846\u67b6\u8bf7\u6c42\u961f\u5217\u6536\u96c6\u7528\u6237\u8bf7\u6c42\uff0c\u7136\u540e\u8fdb\u884c\u4e8c\u6b21\u6279\u91cf\u5904\u7406\u548c\u4f18\u5148\u7ea7\u8c03\u6574\u3002\u63a5\u7740\uff0c\u8fd9\u4e9b\u8bf7\u6c42\u88ab\u9001\u5165 AR \u6a21\u578b\u8fdb\u884c\u8fde\u7eed\u6027\u7684\u63a8\u7406\u3002\u5b8c\u6210\u63a8\u7406\u540e\uff0c\u5f02\u6b65\u5f0f\u5730\u8fd4\u56de\u7ed3\u679c\uff0c\u5e76\u5c06\u5176\u653e\u5165 NAR \u961f\u5217\u4e2d\u7b49\u5f85\u8fdb\u4e00\u6b65\u5904\u7406\u3002\u6700\u540e\uff0c\u5c06\u5904\u7406\u597d\u7684\u7ed3\u679c\u53d1\u9001\u56de\u7528\u6237\u7aef\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s2.51cto.com\/oss\/202411\/11\/a623ec087a2e93327ac181e52c2283f16e8cfe.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<h3>5.\u00a0\u63a8\u7406\u52a0\u901f\u6548\u679c<\/h3>\n<p>\u5b9e\u65f6\u7387\u4ece 2.09 \u4f18\u5316\u5230 0.085\uff0c\u541e\u5410\u53ef\u8fbe\u5230 1 \u79d2 2500tokens\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s7.51cto.com\/oss\/202411\/11\/68e311b06e14bd82bf2968c5320d8ebafb1500.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u56db\u3001\u672a\u6765\u5c55\u671b<\/h2>\n<p>\u5728\u672a\u6765\u7684\u7814\u7a76\u4e0e\u5f00\u53d1\u4e2d\uff0c\u6211\u4eec\u8ba1\u5212\u5b9e\u65bd\u4ee5\u4e0b\u51e0\u9879\u5173\u952e\u4e3e\u63aa\u4ee5\u8fdb\u4e00\u6b65\u4f18\u5316\u8bed\u97f3\u5408\u6210\u6280\u672f\u3002<\/p>\n<ul data-id=\"ua73dd4b-ih24E9SO\">\n<li data-id=\"ld70c578-q7YNNB6G\">\u5f15\u5165\u6295\u673a\u91c7\u6837\u6280\u672f\uff1a\u6211\u4eec\u5c06\u63a2\u7d22\u5c06\u6295\u673a\u91c7\u6837\u65b9\u6cd5\u5e94\u7528\u4e8e\u5927\u578b\u8bed\u97f3\u5408\u6210\u6a21\u578b\u7684\u53ef\u80fd\u6027\u3002\u901a\u8fc7\u8fd9\u79cd\u6280\u672f\uff0c\u53ef\u4ee5\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u63d0\u524d\u9884\u6d4b\u5e76\u9009\u62e9\u6700\u6709\u53ef\u80fd\u7684\u8f93\u51fa\u8def\u5f84\uff0c\u4ece\u800c\u52a0\u901f\u97f3\u9891\u5408\u6210\u8fc7\u7a0b\uff0c\u5e76\u53ef\u80fd\u63d0\u9ad8\u6700\u7ec8\u5408\u6210\u97f3\u9891\u7684\u8d28\u91cf\u3002<\/li>\n<li data-id=\"ld70c578-PnoU1qIY\">\u6539\u9020\u975e\u81ea\u56de\u5f52\uff08NAR\uff09\u6a21\u578b\u4e3a\u6d41\u5f0f\u8f93\u51fa\u67b6\u6784\uff1a\u4e3a\u4e86\u5b9e\u73b0\u66f4\u4f4e\u5ef6\u8fdf\u548c\u66f4\u9ad8\u6548\u7684\u97f3\u9891\u751f\u6210\uff0c\u6211\u4eec\u6253\u7b97\u5bf9\u73b0\u6709\u7684 NAR \u6a21\u578b\u8fdb\u884c\u8c03\u6574\uff0c\u4f7f\u5176\u652f\u6301\u6d41\u5f0f\u8f93\u51fa\u6a21\u5f0f\u3002\u8fd9\u6837\uff0c\u6a21\u578b\u80fd\u591f\u4e00\u8fb9\u751f\u6210\u97f3\u9891\u6570\u636e\u4e00\u8fb9\u5b9e\u65f6\u8f93\u51fa\uff0c\u6781\u5927\u5730\u63d0\u5347\u4e86\u7528\u6237\u4f53\u9a8c\uff0c\u7279\u522b\u662f\u5728\u9700\u8981\u5373\u65f6\u54cd\u5e94\u7684\u5e94\u7528\u573a\u666f\u4e2d\u3002<\/li>\n<li data-id=\"ld70c578-vfkBCD9U\">\u7814\u7a76\u66f4\u591a\u975e Transformer \u7f51\u7edc\u67b6\u6784\uff1a\u867d\u7136 Transformer \u67b6\u6784\u5728\u8bb8\u591a\u65b9\u9762\u8868\u73b0\u4f18\u5f02\uff0c\u4f46\u6211\u4eec\u8ba4\u4e3a\u63a2\u7d22\u5176\u4ed6\u7c7b\u578b\u7684\u795e\u7ecf\u7f51\u7edc\u7ed3\u6784\u540c\u6837\u91cd\u8981\u3002\u8fd9\u5305\u62ec\u4f46\u4e0d\u9650\u4e8e\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff09\u3001\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08RNN\uff09\u53ca\u5176\u53d8\u4f53\u7b49\u3002\u6211\u4eec\u7684\u76ee\u6807\u662f\u53d1\u73b0\u90a3\u4e9b\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u5177\u6709\u4f18\u52bf\u4e14\u80fd\u63d0\u4f9b\u66f4\u9ad8\u6548\u7387\u6216\u66f4\u597d\u6027\u80fd\u7684\u65b0\u67b6\u6784\u3002<\/li>\n<\/ul>\n<\/div>\n<\/div>\n<p>\u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_23012\" class=\"pvc_stats total_only  \" data-element-id=\"23012\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u968f\u7740\u4eba\u5de5\u667a\u80fd\u6280\u672f\u7684\u4e0d\u65ad\u8fdb\u6b65\uff0c\u8bed\u97f3\u5408\u6210\u6280\u672f\u5728\u6e38\u620f\u548c\u5a31\u4e50\u9886\u57df\u626e\u6f14\u7740\u8d8a\u6765\u8d8a\u91cd\u8981\u7684\u89d2\u8272\u3002\u672c\u6b21\u5206\u4eab\u9898\u76ee\u4e3a\u201c\u817e\u8baf\u6e38\u620f\u77e5\u51e0\u8bed [&hellip;]<\/p>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_23012\" class=\"pvc_stats total_only  \" data-element-id=\"23012\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"author":56,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23,20,80],"tags":[113,1423,1424],"class_list":["post-23012","post","type-post","status-publish","format-standard","hentry","category-23","category-20","category-80","tag-113","tag-1423","tag-1424"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/23012","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=23012"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/23012\/revisions"}],"predecessor-version":[{"id":23013,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/23012\/revisions\/23013"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=23012"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=23012"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=23012"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}