{"id":37444,"date":"2025-01-28T19:00:29","date_gmt":"2025-01-28T11:00:29","guid":{"rendered":"https:\/\/17aitech.com\/?p=37444"},"modified":"2025-01-28T19:00:29","modified_gmt":"2025-01-28T11:00:29","slug":"4%e6%af%94%e7%89%b9%e9%87%8f%e5%8c%96%e4%b8%89%e5%80%8d%e5%8a%a0%e9%80%9f%e4%b8%8d%e6%8e%89%e7%82%b9%ef%bc%81%e6%b8%85%e5%8d%8e%e5%8d%b3%e6%8f%92%e5%8d%b3%e7%94%a8%e7%9a%84sageattention%e8%bf%8e","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=37444","title":{"rendered":"4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389\u70b9\uff01\u6e05\u534e\u5373\u63d2\u5373\u7528\u7684SageAttention\u8fce\u6765\u5347\u7ea7"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-12-26-7\" target=\"_blank\">4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389\u70b9\uff01\u6e05\u534e\u5373\u63d2\u5373\u7528\u7684SageAttention\u8fce\u6765\u5347\u7ea7<\/a><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d8de101d64a2e2e8968bb32fbed9c860.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d8de101d64a2e2e8968bb32fbed9c860.png\"><\/a><\/section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"166\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>AIxiv\u4e13\u680f\u662f\u673a\u5668\u4e4b\u5fc3\u53d1\u5e03\u5b66\u672f\u3001\u6280\u672f\u5185\u5bb9\u7684\u680f\u76ee\u3002\u8fc7\u53bb\u6570\u5e74\uff0c\u673a\u5668\u4e4b\u5fc3AIxiv\u4e13\u680f\u63a5\u6536\u62a5\u9053\u4e862000\u591a\u7bc7\u5185\u5bb9\uff0c\u8986\u76d6\u5168\u7403\u5404\u5927\u9ad8\u6821\u4e0e\u4f01\u4e1a\u7684\u9876\u7ea7\u5b9e\u9a8c\u5ba4\uff0c\u6709\u6548\u4fc3\u8fdb\u4e86\u5b66\u672f\u4ea4\u6d41\u4e0e\u4f20\u64ad\u3002\u5982\u679c\u60a8\u6709\u4f18\u79c0\u7684\u5de5\u4f5c\u60f3\u8981\u5206\u4eab\uff0c\u6b22\u8fce\u6295\u7a3f\u6216\u8005\u8054\u7cfb\u62a5\u9053\u3002\u6295\u7a3f\u90ae\u7bb1\uff1aliyazhou@jiqizhixin.com\uff1bzhaoyunfeng@jiqizhixin.com<\/section>\n<\/section>\n<\/blockquote>\n<section><\/section>\n<section><strong>\u8bba\u6587\u5171\u540c\u7b2c\u4e00\u4f5c\u8005\u5f20\u91d1\u6d9b\u3001\u9ec4\u6d69\u5cf0\u5206\u522b\u6765\u81ea\u6e05\u534e\u5927\u5b66\u8ba1\u7b97\u673a\u7cfb\u548c\u4ea4\u53c9\u4fe1\u606f\u7814\u7a76\u9662\uff0c\u8bba\u6587\u901a\u8baf\u4f5c\u8005\u9648\u952e\u98de\u526f\u6559\u6388\u53ca\u5176\u4ed6\u5408\u4f5c\u4f5c\u8005\u5747\u6765\u81ea\u6e05\u534e\u5927\u5b66\u8ba1\u7b97\u673a\u7cfb\u3002<\/strong><\/section>\n<section><\/section>\n<section>\u5927\u6a21\u578b\u4e2d\uff0c\u7ebf\u6027\u5c42\u7684\u4f4e\u6bd4\u7279\u91cf\u5316\u5df2\u7ecf\u9010\u6b65\u843d\u5730\u3002\u7136\u800c\uff0c\u5bf9\u4e8e\u6ce8\u610f\u529b\u6a21\u5757\uff0c\u76ee\u524d\u51e0\u4e4e\u5404\u4e2a\u6a21\u578b\u90fd\u8fd8\u5728\u7528\u9ad8\u7cbe\u5ea6\uff08\u4f8b\u5982 FP16 \u6216 FP32\uff09\u7684\u6ce8\u610f\u529b\u8fd0\u7b97\u8fdb\u884c\u8bad\u7ec3\u548c\u63a8\u7406\u3002\u5e76\u4e14\uff0c\u968f\u7740\u5927\u578b\u6a21\u578b\u9700\u8981\u5904\u7406\u7684\u5e8f\u5217\u957f\u5ea6\u4e0d\u65ad\u589e\u52a0\uff0cAttention\uff08\u6ce8\u610f\u529b\u8fd0\u7b97\uff09\u7684\u65f6\u95f4\u5f00\u9500\u9010\u6e10\u6210\u4e3a\u4e3b\u8981\u5f00\u9500\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u524d\uff0c\u6e05\u534e\u5927\u5b66\u9648\u952e\u98de\u56e2\u961f\u63d0\u51fa\u7684 8-Bit \u7684\u5373\u63d2\u5373\u7528 Attention\uff08<a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"https:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650939511&amp;idx=3&amp;sn=dabddded9914c9219df01f93576b1930&amp;scene=21#wechat_redirect\" target=\"_blank\">SageAttention<\/a>\uff09\uff0c\u5c06 Attention \u4e2d\u7684 QK^T \u91cf\u5316\u81f3 INT8\uff0c\u5c06 PV \u4fdd\u6301\u4e3a FP16 \u7cbe\u5ea6\u5e76\u4f7f\u7528 FP16 \u7cbe\u5ea6\u7684\u77e9\u9635\u4e58\u6cd5\u7d2f\u52a0\u5668\uff0c\u540c\u65f6\u63d0\u51fa Smooth K \u6280\u672f\u4fdd\u6301\u4e86\u91cf\u5316 Attention \u7684\u7cbe\u5ea6\uff0c\u5b9e\u73b0\u4e86 2 \u500d\u52a0\u901f\u4e8e FlashAttention2\uff0c\u4e14\u5728\u5404\u7c7b\u5927\u6a21\u578b\u4e0a\u5747\u4fdd\u6301\u4e86\u7aef\u5230\u7aef\u7684\u7cbe\u5ea6\u8868\u73b0\u3002<\/section>\n<section><\/section>\n<section>\u76ee\u524d\uff0cSageAttention \u5df2\u7ecf\u88ab\u4e1a\u754c\u53ca\u793e\u533a\u5e7f\u6cdb\u5730\u4f7f\u7528\u4e8e\u5404\u79cd\u5f00\u6e90\u53ca\u5546\u4e1a\u5927\u6a21\u578b\u4e2d\uff0c\u6bd4\u5982 CogvideoX\u3001Mochi\u3001Flux\u3001Llama3\u3001Qwen \u7b49\u3002<\/section>\n<section><\/section>\n<section>\u8fd1\u65e5\uff0c\u9648\u952e\u98de\u56e2\u961f\u8fdb\u4e00\u6b65\u63d0\u51fa\u4e86 4-Bit \u7684\u5373\u63d2\u5373\u7528 Attention\uff08SageAttention2\uff09\uff0c\u76f8\u8f83\u4e8e FlashAttention2 \u548c xformers \u5206\u522b\u5b9e\u73b0\u4e86 <strong>3\u00a0<\/strong>\u500d\u4ee5\u53ca<strong>\u00a04.5<\/strong> \u500d\u7684<strong>\u5373\u63d2\u5373\u7528<\/strong>\u7684\u63a8\u7406\u52a0\u901f\uff0c\u4e14\u5728\u89c6\u9891\u3001\u56fe\u50cf\u3001\u6587\u672c\u751f\u6210\u7b49\u5927\u6a21\u578b\u4e0a<strong>\u5747\u4fdd\u6301\u4e86\u7aef\u5230\u7aef\u7684\u7cbe\u5ea6\u8868\u73b0<\/strong>\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-cb9d460ed78bc416b45b07a690a22422.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-cb9d460ed78bc416b45b07a690a22422.png\"><\/a><\/section>\n<ul>\n<li>\n<p>\u8bba\u6587\u6807\u9898\uff1aSageAttention2: Efficient Attention with Thorough Outlier Smoothing and Per-thread INT4 Quantization<\/p>\n<\/li>\n<li>\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2411.10958<\/p>\n<\/li>\n<li>\n<p>\u5f00\u6e90\u4ee3\u7801\uff1ahttps:\/\/github.com\/thu-ml\/SageAttention<\/p>\n<\/li>\n<\/ul>\n<section><strong>\u5373\u63d2\u5373\u7528\u4e3e\u4f8b<\/strong><\/section>\n<section><\/section>\n<section>SageAttention2 \u5b9e\u73b0\u4e86\u9ad8\u6548\u7684 Attention \u7b97\u5b50\uff0c\u53ef\u4ee5\u5b9e\u73b0\u5373\u63d2\u5373\u7528\u7684\u63a8\u7406\u52a0\u901f\u3002\u8f93\u5165\u4efb\u610f Q, K, V \u77e9\u9635\uff0cSageAttention2 \u53ef\u4ee5\u5feb\u901f\u8fd4\u56de Attention Output (O)\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f2d9a0afd4ef987f79e75d5e8a6152e7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f2d9a0afd4ef987f79e75d5e8a6152e7.png\"><\/a><\/section>\n<section>\u5177\u4f53\u6765\u8bf4\uff0cSageAttention2 \u4f7f\u7528\u8d77\u6765\u5f88\u65b9\u4fbf\uff0c\u514b\u9686\u4ed3\u5e93\uff08git clone https:\/\/github.com\/thu-ml\/SageAttention\uff09\u5e76\u6267\u884c python setup.py install \u540e\uff0c\u53ea\u9700\u4e00\u884c\u4ee3\u7801\u4fbf\u53ef\u4ee5\u5f97\u5230 Attention \u7684\u8f93\u51fa\uff0c\u53ef\u4ee5\u4f7f\u7528\u8be5\u63a5\u53e3\u65b9\u4fbf\u5730\u66ff\u6362\u4efb\u610f\u6a21\u578b\u4e2d\u7684 Attention \u51fd\u6570\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-909110bc2cae8882271ab96e265f8bb9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-909110bc2cae8882271ab96e265f8bb9.png\"><\/a><\/section>\n<section>\u6548\u679c\u4e0a\uff0c\u4ee5\u5f00\u6e90\u89c6\u9891\u751f\u6210\u6a21\u578b CogvideoX-1.5-5B \u4e3a\u4f8b\uff0c\u4f7f\u7528 SageAttention2 \u53ef\u4ee5\u7aef\u5230\u7aef\u52a0\u901f\u00a0<strong>1.8\u00a0<\/strong>\u500d\uff0c\u4e14\u751f\u6210\u7684\u89c6\u9891\u65e0\u635f\uff1a<a href=\"https:\/\/mp.weixin.qq.com\/s\/Zl-rbVn_mqujDVrbBcCShw?token=2132360016&amp;lang=zh_CN\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-6a3c42d48455926b0e90458701446cf5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-6a3c42d48455926b0e90458701446cf5.png\"><\/a><\/a><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u4f7f\u7528\u5168\u7cbe\u5ea6 Attention<\/sup><\/em><sup>\u00a0<\/sup><a href=\"https:\/\/mp.weixin.qq.com\/s\/Zl-rbVn_mqujDVrbBcCShw?token=2132360016&amp;lang=zh_CN\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-9c4285b6a34944d34359db99e543945a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-9c4285b6a34944d34359db99e543945a.png\"><\/a><\/a><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u4f7f\u7528 SageAttention2<\/sup><\/em><sup>\u00a0<\/sup><\/section>\n<section><\/section>\n<section>\u66f4\u91cd\u8981\u7684\u662f\uff0cSageAttention2 \u63d0\u4f9b\u4e86\u6bd4 SageAttention \u66f4\u5e7f\u6cdb\u7684\u786c\u4ef6\u652f\u6301\u3002\u9664\u4e86\u5728 RTX 4090 \u4e0a\u53ef\u4ee5 3 \u500d\u52a0\u901f\u4e8e FlashAttention \u5916\uff0c\u5728 L20\u3001L40\u3001L40S \u53ef\u4ee5\u5b9e\u73b0 2 \u500d\u7684\u52a0\u901f\uff0c\u5728 A100\u3001A800\u3001A6000 \u4e0a\u53ef\u4ee5\u5b9e\u73b0 1.45-1.6 \u500d\u7684\u52a0\u901f\uff08\u57fa\u4e8e SageAttention\uff09\u3002<\/section>\n<section><\/section>\n<section>\u63a5\u4e0b\u6765\uff0c\u7814\u7a76\u56e2\u961f\u5c06\u4ece\u524d\u8a00\u3001\u6311\u6218\u3001\u65b9\u6cd5\u4ee5\u53ca\u5b9e\u9a8c\u6548\u679c\u56db\u4e2a\u65b9\u9762\u4ecb\u7ecd SageAttention2\uff08\u603b\u4f53\u6d41\u7a0b\u56fe\u5982\u4e0b\u56fe\uff09\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4c720aa28afa357ad215f678bdd9e321.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4c720aa28afa357ad215f678bdd9e321.png\"><\/a><\/section>\n<section><strong>\u524d\u8a00<\/strong><\/section>\n<section><\/section>\n<section>\u968f\u7740\u5927\u6a21\u578b\u9700\u8981\u5904\u7406\u7684\u5e8f\u5217\u957f\u5ea6\u8d8a\u6765\u8d8a\u957f\uff0cAttention \u7684\u901f\u5ea6\u4f18\u5316\u53d8\u5f97\u8d8a\u6765\u8d8a\u91cd\u8981\u3002\u4e0b\u56fe\u5c55\u793a\u4e86\u4e00\u4e2a\u6807\u51c6\u7684 Transformer \u6a21\u578b\u4e2d\u5404\u8fd0\u7b97\u7684\u65f6\u95f4\u5360\u6bd4\u968f\u5e8f\u5217\u957f\u5ea6\u7684\u53d8\u5316\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-3897c797cc9430f07662e21f732be616.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-3897c797cc9430f07662e21f732be616.png\"><\/a><\/section>\n<section>\u4e3a\u4e86\u65b9\u4fbf\u6307\u4ee3\u6ce8\u610f\u529b\u8fd0\u7b97\u4e2d\u7684\u77e9\u9635\uff0c\u6211\u4eec\u5148\u56de\u987e\u4e00\u4e0b\u6ce8\u610f\u529b\u7684\u8ba1\u7b97\u516c\u5f0f\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4abe2f2f9d70199c2350a8f199ea8dfc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4abe2f2f9d70199c2350a8f199ea8dfc.png\"><\/a><\/section>\n<section>\u5c3d\u7ba1 SageAttention \u63d0\u51fa\u5c06 Q\uff0cK \u91cf\u5316\u81f3 INT8\uff0c\u5c06 P\uff0cV \u4fdd\u6301 FP16 \u7cbe\u5ea6\u4e14\u91c7\u7528 FP16 \u7684\u77e9\u9635\u4e58\u6cd5\u7d2f\u52a0\u5668\u6765\u52a0\u5feb Attention \u7684\u901f\u5ea6\u3002\u7136\u800c\uff0c\u8fd9\u6837\u505a\u7684\u7f3a\u70b9\u662f\uff1a1\uff09INT8 \u7684\u77e9\u9635\u4e58\u6cd5\u53ea\u8fbe\u5230\u4e86\u4e00\u534a\u7684 INT4 \u77e9\u9635\u4e58\u6cd5\u7684\u901f\u5ea6\uff0c2\uff09\u4f7f\u7528 FP16 \u7684\u4e58\u6cd5\u7d2f\u52a0\u5668\u7684 FP16 \u7684\u77e9\u9635\u4e58\u6cd5\u7684\u52a0\u901f\u53ea\u5728 RTX4090 \u548c RTX3090 \u663e\u5361\u4e0a\u6709\u6548\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u514b\u670d\u4e0a\u8ff0\u7f3a\u70b9\uff0cSageAttention2 \u63d0\u51fa\u5c06 Q, K \u91cf\u5316\u81f3 INT4\uff0c\u5e76\u5c06 P, V \u91cf\u5316\u81f3 FP8 \u6765\u52a0\u901f Attention\u3002\u7136\u800c\uff0c\u8fd9\u6837\u505a\u7684\u6311\u6218\u662f\u5f88\u5927\u7684\u3002<\/section>\n<section><\/section>\n<section><strong>4-Bit \u6ce8\u610f\u529b\u91cf\u5316\u6709\u4ec0\u4e48\u95ee\u9898\uff1f<\/strong><\/section>\n<section><\/section>\n<section>\u7814\u7a76\u56e2\u961f\u53d1\u73b0\u76f4\u63a5\u5c06\u6ce8\u610f\u529b\u8fd0\u7b97\u4e2d\u7684 Q, K \u91cf\u5316\u4e3a INT4 \u540e\u5c06\u4f1a\u5bfc\u81f4\u5728\u51e0\u4e4e\u6240\u6709\u6a21\u578b\u548c\u4efb\u52a1\u4e0a\u90fd\u4f1a\u5f97\u5230\u6781\u5dee\u7684\u7ed3\u679c\uff0c\u4f8b\u5982\uff0c\u5728 CogVideoX \u6587\u751f\u89c6\u9891\u6a21\u578b\u4e2d\uff0c\u4f1a\u5f97\u5230\u5b8c\u5168\u6a21\u7cca\u7684\u89c6\u9891\uff1bLlama2-7B \u8fdb\u884c\u56db\u9009\u4e00\u9009\u62e9\u9898\u4efb\u52a1\u4e0a\u5f97\u5230 25% \u7684\u51c6\u786e\u7387\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-769dce4e92e7609fd90ed5254c2273e2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-769dce4e92e7609fd90ed5254c2273e2.png\"><\/a><\/section>\n<section>\u7ecf\u8fc7\u4ed4\u7ec6\u5206\u6790\u540e\uff0c\u7814\u7a76\u56e2\u961f\u53d1\u73b0\u4e3b\u8981\u662f\u4e24\u4e2a\u539f\u56e0\u5bfc\u81f4\u4e86\u91cf\u5316\u6ce8\u610f\u529b\u7684\u4e0d\u51c6\u786e\uff1a<\/section>\n<section><\/section>\n<section>\uff081\uff09INT4 \u7684\u6570\u503c\u8303\u56f4\u76f8\u6bd4 INT8 \u975e\u5e38\u5c0f\uff0c\u5bfc\u81f4\u5176\u91cf\u5316\u8bef\u5dee\u5728 Q\uff0cK \u77e9\u9635\u4e2d\u51fa\u73b0\u4e00\u4e9b\u5f02\u5e38\u503c\u65f6\u4f1a\u53d8\u5f97\u5341\u5206\u660e\u663e\uff0c\u6070\u597d\u5927\u591a\u6a21\u578b\u90fd\u5728 Q, K \u4e2d\u8868\u73b0\u51fa\u6765\u4e86\u8f83\u5927\u7684\u901a\u9053\u7ef4\u5ea6\u7684\u5f02\u5e38\u503c\u3002\u8fd9\u6781\u5927\u524a\u51cf\u4e86 QK^\u22a4\u77e9\u9635\u4e58\u6cd5\u7684\u7cbe\u5ea6\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-6f6ec5950d6d4fdc3051ab002da9c265.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-6f6ec5950d6d4fdc3051ab002da9c265.png\"><\/a><\/section>\n<section>\uff082\uff09\u7814\u7a76\u56e2\u961f\u53d1\u73b0 Nvidia \u7684\u663e\u5361\u4e0a\uff0cFP8 \u7684\u77e9\u9635\u4e58\u6cd5\u6307\u4ee4 (mma.f32.f8.f8.f32) \u7684\u4e58\u6cd5\u7d2f\u52a0\u5668\u5e76\u4e0d\u662f\u5b98\u65b9\u5ba3\u79f0\u7684 FP32 \u7cbe\u5ea6\uff0c\u800c\u662f\u53ea\u6709 FP22 \u7cbe\u5ea6\uff0c\u8fd9\u5bfc\u81f4\u4e86 PV \u77e9\u9635\u4e58\u6cd5\u51fa\u73b0\u8f83\u5927\u7684\u7d2f\u52a0\u8bef\u5dee\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-c21be5b1f476021b70b8c89e667119f1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-c21be5b1f476021b70b8c89e667119f1.png\"><\/a><\/section>\n<section><strong>\u6280\u672f\u65b9\u6848<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u7684\u4e24\u4e2a\u6311\u6218\uff0c\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86\u5bf9\u5e94\u7684\u89e3\u51b3\u529e\u6cd5\u3002<\/section>\n<section><\/section>\n<section>\uff081\uff09\u4fdd\u7559 SageAttention \u4e2d\u5bf9 K \u8fdb\u884c\u5e73\u6ed1\u5904\u7406\u7684\u540c\u65f6\uff0c\u63d0\u51fa\u5bf9 Q \u8fdb\u884c\u5e73\u6ed1\u5904\u7406\uff1aQ \u2013 mean (Q)\u3002\u5176\u4e2d mean (Q) \u662f\u6cbf\u7740\u901a\u9053\u7ef4\u5ea6\u7684\u5e73\u5747\u503c\u5411\u91cf\u3002\u5b8c\u6210\u8be5\u5e73\u6ed1\u64cd\u4f5c\u540e\u9700\u8981\u5728 Attention \u8ba1\u7b97\u8fc7\u7a0b\u4e2d\u5c06 mean (Q) \u548c K^T \u7684\u5411\u91cf\u4e0e\u77e9\u9635\u4e58\u6cd5\u7684\u7ed3\u679c\u8865\u507f\u5230 S \u4e2d\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f7ab74b31db222773779c02053d20abd.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f7ab74b31db222773779c02053d20abd.png\"><\/a><\/section>\n<section>\u8fd9\u4f7f\u5f97\u76f8\u6bd4\u76f4\u63a5\u91cf\u5316 Q, K \u81f3 INT4 \u7684\u51c6\u786e\u5ea6\u6709\u8d28\u7684\u6539\u53d8\uff0c\u5982\u4e0b\u8868\u5c55\u793a\u4e86\u5bf9\u6bd4\u4e86\u8be5\u65b9\u6cd5\u548c\u76f4\u63a5\u91cf\u5316 Q, K \u81f3 INT4 \u5728 Cogvideo \u548c Llama3.1 \u4e0a\u7684\u7aef\u5230\u7aef\u8868\u73b0\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-7434ac3654c804121a2009caa83dc9c0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-7434ac3654c804121a2009caa83dc9c0.png\"><\/a><\/section>\n<section>\u77e9\u9635 Q \u5e73\u6ed1\u524d\u540e\u7684\u6570\u636e\u5206\u5e03\u53ef\u89c6\u5316\u7684\u7ed3\u679c\u5982\u4e0b\uff0c\u53ef\u4ee5\u53d1\u73b0\u5e73\u6ed1\u540e\u7684 Q \u5bf9 INT4 \u6570\u636e\u8303\u56f4\u7684\u5229\u7528\u5ea6\u66f4\u9ad8\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-db6d98a8ecdd0ad906ac61df5efd6163.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-db6d98a8ecdd0ad906ac61df5efd6163.png\"><\/a><\/section>\n<section>\uff082\uff09\u5bf9 Q, K \u8fdb\u884c Per-thread \u91cf\u5316\u3002\u5bf9\u4e8e\u77e9\u9635 Q, K\uff0cSageAttention2 \u91c7\u7528\u4e86\u6839\u636e mma \u6307\u4ee4\u5bf9\u77e9\u9635\u5185\u5b58\u6392\u5e03\u7684\u8981\u6c42\uff0c\u5bf9 Q\uff0cK \u4e2d\u7684 Token \u6309\u7167 GPU \u7ebf\u7a0b\u8fdb\u884c\u5206\u7ec4\uff0c\u4f7f\u91cf\u5316\u7c92\u5ea6\u6bd4 SageAttention \u4e2d\u7684 per-block \u7ec6\u5316 16 \u500d\uff0c\u6781\u5927\u63d0\u9ad8\u4e86 4Bit \u7684 QK^\u22a4\u4e58\u6cd5\u51c6\u786e\u5ea6\u7684\u540c\u65f6\u4e0d\u5f15\u5165\u4efb\u4f55\u989d\u5916\u5f00\u9500\u3002<\/section>\n<section><\/section>\n<section>\u5177\u4f53\u6765\u8bf4\uff0c\u5728 SageAttention \u4e2d\uff0c\u6bcf\u4e2a Q \u7684\u5757\u5c06\u88ab\u5212\u5206\u4e3a c_w \u4e2a\u6bb5\uff0c\u7531 GPU \u6d41\u5904\u7406\u5668\uff08SM\uff09\u4e2d\u7684 c_w \u4e2a GPU warp \u5904\u7406\u3002\u7136\u540e\uff0c\u6bcf\u4e2a\u5305\u542b 32 \u4e2a\u7ebf\u7a0b\u7684 warp \u4f1a\u4f7f\u7528 NVIDIA \u7684 mma.m16n8k64 PTX \u6307\u4ee4\u6765\u6267\u884c QK^\u22a4\u8fd0\u7b97\u3002\u6839\u636e\u8fd9\u4e00\u6307\u4ee4\u7684\u5e03\u5c40\u8981\u6c42\uff0c\u7814\u7a76\u56e2\u961f\u53d1\u73b0\u4e00\u4e2a warp \u5185\u7684 Q [8\u00d7(n%8)] \u53ef\u4ee5\u5171\u7528\u4e00\u4e2a\u91cf\u5316\u7f29\u653e\u53c2\u6570\uff0c\u800c\u4e00\u4e2a warp \u5185\u7684 K [8\u00d7(n%8)] \u548c K [8\u00d7(n%8+1)] \u4e5f\u53ef\u4ee5\u5171\u7528\u4e00\u4e2a\u91cf\u5316\u7f29\u653e\u53c2\u6570\uff0c\u5176\u4e2d n \u662f token \u7d22\u5f15\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u79cd\u91cf\u5316\u65b9\u6cd5\u66f4\u4e3a\u7ec6\u81f4\u4e14\u4e0d\u589e\u52a0\u989d\u5916\u5f00\u9500\u3002\u8fd9\u662f\u56e0\u4e3a\u5b83\u6839\u636e MMA \u6307\u4ee4\u7684\u5e03\u5c40\u5c06\u4e0d\u540c\u7684 GPU \u7ebf\u7a0b\u5206\u914d\u5230\u4e0d\u540c\u7684\u91cf\u5316 Token \u7ec4\uff0c\u6bcf\u4e2a\u7ebf\u7a0b\u53ea\u5bf9\u5e94\u4e00\u4e2a\u91cf\u5316\u7f29\u653e\u53c2\u6570\u8fdb\u884c\u53cd\u91cf\u5316\u3002\u800c\u975e Per-token \u91cf\u5316\u90a3\u6837\uff0c\u6bcf\u4e2a\u7ebf\u7a0b\u5bf9\u5e94\u591a\u4e2a\u91cf\u5316\u7f29\u653e\u53c2\u6570\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-2035bd3821afe83646913cef748f473d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-2035bd3821afe83646913cef748f473d.png\"><\/a><\/section>\n<section>\u5982\u4e0b\u8868\u6240\u793a\uff0c\u53ef\u4ee5\u53d1\u73b0 per-thread \u91cf\u5316\u7684\u51c6\u786e\u5ea6\u6bd4 SageAttention \u4e2d\u91c7\u7528\u7684 per-block \u91cf\u5316\u9ad8\u5f97\u591a\uff0c\u51c6\u786e\u5ea6\u548c per-token \u91cf\u5316\u51e0\u4e4e\u6ca1\u6709\u5dee\u522b\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-16ffbf54e2d0caa36aa6ad67fa7a1a70.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-16ffbf54e2d0caa36aa6ad67fa7a1a70.png\"><\/a><\/section>\n<section>\uff083\uff09\u5bf9 FP8 \u7684 PV \u77e9\u9635\u4e58\u6cd5\u91c7\u7528 FP32 \u7684\u5bc4\u5b58\u5668\u5c06\u6bcf\u6b21 FlashAttention \u5206\u5757\u7c92\u5ea6\u7684 PV \u7684 FP22 \u7684\u4e58\u6cd5\u7ed3\u679c\u7d2f\u52a0\u8d77\u6765\u3002\u8fd9\u79cd\u505a\u6cd5\u53ef\u4ee5\u6709\u6548\u5730\u907f\u514d FP22 \u7684\u4e58\u6cd5\u7d2f\u52a0\u5668\u6cbf\u7740\u5e8f\u5217\u957f\u5ea6\u7d2f\u79ef\u8fc7\u591a\u7684\u8bef\u5dee\uff0c\u5c06 FP22 \u7d2f\u52a0\u5668\u5e26\u6765\u7684\u8bef\u5dee\u63a7\u5236\u5728 FlashAttention \u5206\u5757\u7684\u7c92\u5ea6\u4e2d\uff0c\u63d0\u9ad8\u4e86 FP8 \u7684 PV \u4e58\u6cd5\u7684\u51c6\u786e\u5ea6\u3002<\/section>\n<section><\/section>\n<section>\uff084\uff09\u9488\u5bf9 P \u548c V\uff0c\u7814\u7a76\u56e2\u961f\u5bf9\u6bd4\u4e86\u591a\u79cd\u91cf\u5316\u7684\u6570\u636e\u7c7b\u578b\uff0c\u5bf9\u6bd4\u53d1\u73b0\u4f7f\u7528 E4M3 \u6570\u636e\u683c\u5f0f\u7684 FP8 \u7cbe\u5ea6\u6700\u51c6\u786e\uff0c\u57fa\u672c\u63a5\u8fd1\u4e86 FP16 \u7684\u51c6\u786e\u5ea6\u3002\u56e0\u6b64\u91c7\u7528\u5c06 P \u548c V \u91cf\u5316\u81f3 E4M3\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4ff0e79d19a830cf4f23361ec2b26c3f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4ff0e79d19a830cf4f23361ec2b26c3f.png\"><\/a><\/section>\n<section>\u4e0b\u56fe\u5c55\u793a\u4e86 SageAttention2 \u7684\u7b97\u6cd5\u6d41\u7a0b\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-30fb69ef74bb65259c14f887548126b5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-30fb69ef74bb65259c14f887548126b5.png\"><\/a><\/section>\n<section>SageAttention2 \u5171\u5b9e\u73b0\u4e86\u4e24\u79cd Kernel\uff0c\u533a\u522b\u5728\u4e8e\u5bf9 Q, K \u8fdb\u884c INT4 \u91cf\u5316\u8fd8\u662f INT8 \u91cf\u5316\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-2aedff92e54ec1a41fef71fb1d0e52ee.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-2aedff92e54ec1a41fef71fb1d0e52ee.png\"><\/a><\/section>\n<section>\u6b64\u5916\uff0cSageAttention2 \u8fd8\u63d0\u51fa\u4e00\u79cd\u53ef\u9009\u7684\u5bf9\u77e9\u9635 V \u8fdb\u884c\u5e73\u6ed1\u5904\u7406\u7684\u6280\u672f\uff0c\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8 PV \u77e9\u9635\u4e58\u6cd5\u7684\u51c6\u786e\u5ea6\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5f53\u67d0\u4e9b\u6a21\u578b\u4e2d V \u77e9\u9635\u5177\u6709\u901a\u9053\u7ef4\u5ea6\u7684\u504f\u79fb\u65f6\uff0c\u53ef\u4ee5\u5c06 V \u51cf\u53bb\u5176\u901a\u9053\u7ef4\u5ea6\u7684\u5e73\u5747\u503c mean (V) \u6765\u53bb\u9664\u504f\u79fb\uff0c\u4e4b\u540e\u8fdb\u884c\u6b63\u5e38\u7684\u91cf\u5316 Attention \u8fd0\u7b97\u3002\u53ea\u9700\u8981\u5bf9\u6700\u7ec8 Attention \u7684 Output \u52a0\u4e0a mean (V) \u5373\u53ef\u4fdd\u6301\u8ba1\u7b97\u7684\u6b63\u786e\u6027\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-036dca8e0843f158b64b7c4509603f12.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-036dca8e0843f158b64b7c4509603f12.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-6514079076f3ac8e528de598745233b3.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-6514079076f3ac8e528de598745233b3.png\"><\/a><\/section>\n<section>\u8fd9\u79cd\u505a\u6cd5\u53ef\u4ee5\u63d0\u5347\u51c6\u786e\u5ea6\u7684\u539f\u56e0\u5982\u4e0b\u56fe\u6240\u793a\u3002\u5728 FP22 \u7684\u8868\u793a\u8303\u56f4\u5185\uff0c\u6570\u503c\u8d8a\u5927\uff0c\u76f8\u6bd4 FP32 \u7684\u8bef\u5dee\u8d8a\u5927\u3002\u800c P \u7684\u8303\u56f4\u662f 0\uff5e1 \u4e4b\u95f4\uff0c\u90a3\u4e48\u5f53 V \u77e9\u9635\u7684\u5217\u6709\u8f83\u5927\u7684\u6570\u503c\u504f\u79fb\u65f6\uff0cPV \u7684 FP22 \u7d2f\u52a0\u5668\u7684\u7cbe\u5ea6\u5c31\u8d8a\u5dee\uff0c\u901a\u8fc7\u5e73\u6ed1 V \u53bb\u9664\u504f\u79fb\u540e\uff0c\u5c31\u53ef\u4ee5\u52a0\u5f3a PV \u77e9\u9635\u4e58\u6cd5\u7684\u51c6\u786e\u5ea6\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-c81bb05bfecbbb43d655440e6267c0e0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-c81bb05bfecbbb43d655440e6267c0e0.png\"><\/a><\/section>\n<section><strong>\u5b9e\u9a8c\u6548\u679c<\/strong><\/section>\n<section><\/section>\n<section>SageAttention \u5b9e\u73b0\u4e86\u5e95\u5c42\u7684 GPU CUDA Kernel\uff0c\u5728\u7b97\u5b50\u901f\u5ea6\u4ee5\u53ca\u5404\u4e2a\u6a21\u578b\u7aef\u5230\u7aef\u51c6\u786e\u5ea6\u4e0a\u90fd\u6709\u5341\u5206\u4e0d\u9519\u7684\u8868\u73b0\u3002<\/section>\n<section><\/section>\n<section>\u5177\u4f53\u6765\u8bf4\uff0c\u7b97\u5b50\u901f\u5ea6\u76f8\u6bd4\u4e8e FlashAttention2 \u548c xformers \u6709\u5927\u7ea6 3 \u500d\u4ee5\u53ca 4.5 \u500d\u7684\u52a0\u901f\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-351450a7f1479907d6b8f4180c811f14.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-351450a7f1479907d6b8f4180c811f14.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-299478342bce7eb46155f75854225135.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-299478342bce7eb46155f75854225135.png\"><\/a><\/section>\n<section>\u7b97\u5b50\u7684\u51c6\u786e\u5ea6\u65b9\u9762\u4e5f\u662f\u6bd4\u5bf9 Q, K \u8fdb\u884c SmoothQuant \u548c Hadamard \u53d8\u6362\u8981\u66f4\u52a0\u51c6\u786e\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-067327a60eaadd259a9dae060c56a0e7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-067327a60eaadd259a9dae060c56a0e7.png\"><\/a><\/section>\n<section>\u5404\u6a21\u578b\u5728\u771f\u5b9e\u573a\u666f\u7684\u7aef\u5230\u7aef\u7cbe\u5ea6\u8868\u73b0\u4e2d\uff0c\u5728\u89c6\u9891\u3001\u56fe\u50cf\u3001\u6587\u672c\u751f\u6210\u7b49\u5927\u6a21\u578b\u4e0a\u5747\u4fdd\u6301\u4e86\u7aef\u5230\u7aef\u7684\u7cbe\u5ea6\u8868\u73b0\uff1a<\/section>\n<section><\/section>\n<section>\u4e0b\u56fe\u662f\u5728 HunyuanVideo \u4e2d\u7684\u53ef\u89c6\u5316\u5b9e\u4f8b\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-fff0ec48507efb21fa90fa84a66b7e09.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-fff0ec48507efb21fa90fa84a66b7e09.png\"><\/a><\/section>\n<section>\u4e0b\u56fe\u662f\u5728 Cogvideo \u4e2d\u7684\u53ef\u89c6\u5316\u5b9e\u4f8b\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-927459bad5289c258659dca01d6bf2c2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-927459bad5289c258659dca01d6bf2c2.png\"><\/a><\/section>\n<section>\u4e0b\u8868\u5c55\u793a\u4e86\u5404\u4e2a\u8bed\u8a00\u3001\u89c6\u9891\u3001\u56fe\u50cf\u751f\u6210\u6a21\u578b\u4e2d SageAttention2 \u7684\u7aef\u5230\u7aef\u7cbe\u5ea6\u8868\u73b0\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-415af23896fed474ba916a20227c3e7f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-415af23896fed474ba916a20227c3e7f.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-6738e202fa44b9775fc054cad440d524.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-6738e202fa44b9775fc054cad440d524.png\"><\/a><\/section>\n<section><strong>\u7aef\u5230\u7aef\u7684\u901f\u5ea6\u8868<\/strong><strong>\u73b0\u4e0a<\/strong>\uff0cSageAttention2 \u4e24\u4e2a Kernel \u7684\u5b9e\u73b0\u5747\u53ef\u4ee5\u6709\u6548\u5730\u5bf9\u957f\u5e8f\u5217\u6a21\u578b\u8fdb\u884c\u52a0\u901f\uff0c\u6bd4\u5982\u53ef\u4ee5\u7aef\u5230\u7aef\u00a0<strong>1.8<\/strong> \u500d\u52a0\u901f CogVideoX1.5-5B\uff0c\u5176\u4ed6\u6a21\u578b\u4e0a\u4e5f\u5747\u6709\u00a0<strong>1.6<\/strong> \u5230<strong>\u00a01.8<\/strong> \u500d\u7684\u63d0\u901f\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f0f22cf07603e61c0b1fe6154e6c2106.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f0f22cf07603e61c0b1fe6154e6c2106.png\"><\/a><\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-12-26-7\" target=\"_blank\">4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389\u70b9\uff01\u6e05\u534e\u5373\u63d2\u5373\u7528\u7684SageAttention\u8fce\u6765\u5347\u7ea7<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389 [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68],"class_list":["post-37444","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389\u70b9\uff01\u6e05\u534e\u5373\u63d2\u5373\u7528\u7684SageAttention\u8fce\u6765\u5347\u7ea7 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=37444\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37444#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37444\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389\u70b9\uff01\u6e05\u534e\u5373\u63d2\u5373\u7528\u7684SageAttention\u8fce\u6765\u5347\u7ea7\",\"datePublished\":\"2025-01-28T11:00:29+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37444\"},\"wordCount\":294,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37444#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-d8de101d64a2e2e8968bb32fbed9c860.png\",\"keywords\":[\"\u5927\u6a21\u578b\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=37444#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37444\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=37444\",\"name\":\"4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389\u70b9\uff01\u6e05\u534e\u5373\u63d2\u5373\u7528\u7684SageAttention\u8fce\u6765\u5347\u7ea7 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37444#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37444#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-d8de101d64a2e2e8968bb32fbed9c860.png\",\"datePublished\":\"2025-01-28T11:00:29+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37444#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=37444\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37444#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-d8de101d64a2e2e8968bb32fbed9c860.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-d8de101d64a2e2e8968bb32fbed9c860.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=37444#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389\u70b9\uff01\u6e05\u534e\u5373\u63d2\u5373\u7528\u7684SageAttention\u8fce\u6765\u5347\u7ea7\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\",\"url\":\"https:\\\/\\\/17aitech.com\\\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/17aitech.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"url\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"contentUrl\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389\u70b9\uff01\u6e05\u534e\u5373\u63d2\u5373\u7528\u7684SageAttention\u8fce\u6765\u5347\u7ea7 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=37444","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=37444#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=37444"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389\u70b9\uff01\u6e05\u534e\u5373\u63d2\u5373\u7528\u7684SageAttention\u8fce\u6765\u5347\u7ea7","datePublished":"2025-01-28T11:00:29+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=37444"},"wordCount":294,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=37444#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d8de101d64a2e2e8968bb32fbed9c860.png","keywords":["\u5927\u6a21\u578b"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=37444#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=37444","url":"https:\/\/17aitech.com\/?p=37444","name":"4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389\u70b9\uff01\u6e05\u534e\u5373\u63d2\u5373\u7528\u7684SageAttention\u8fce\u6765\u5347\u7ea7 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=37444#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=37444#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d8de101d64a2e2e8968bb32fbed9c860.png","datePublished":"2025-01-28T11:00:29+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=37444#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=37444"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=37444#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d8de101d64a2e2e8968bb32fbed9c860.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d8de101d64a2e2e8968bb32fbed9c860.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=37444#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"4\u6bd4\u7279\u91cf\u5316\u4e09\u500d\u52a0\u901f\u4e0d\u6389\u70b9\uff01\u6e05\u534e\u5373\u63d2\u5373\u7528\u7684SageAttention\u8fce\u6765\u5347\u7ea7"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/37444","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=37444"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/37444\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=37444"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=37444"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=37444"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}