{"id":550,"date":"2025-08-26T09:34:10","date_gmt":"2025-08-26T01:34:10","guid":{"rendered":"https:\/\/aitimes.link\/?p=550"},"modified":"2025-08-26T09:34:50","modified_gmt":"2025-08-26T01:34:50","slug":"550","status":"publish","type":"post","link":"https:\/\/aitimes.link\/index.php\/2025\/08\/26\/550\/","title":{"rendered":"LLM \u9ad8\u6548\u67b6\u6784\u7efc\u8ff0"},"content":{"rendered":"\n<p>\u8fd1\u5e74\u6765\uff0c\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u8bed\u8a00\u7406\u89e3\u4e0e\u751f\u6210\u80fd\u529b\uff0c\u63a8\u52a8\u4e86\u6587\u672c\u751f\u6210\u3001\u4ee3\u7801\u751f\u6210\u3001\u95ee\u7b54\u3001\u7ffb\u8bd1\u7b49\u4efb\u52a1\u7684\u7a81\u7834\u3002\u4ee3\u8868\u6027\u6a21\u578b\u5982 GPT\u3001Claude\u3001Gemini\u3001DeepSeek\u3001Qwen \u7b49\uff0c\u5df2\u7ecf\u6df1\u523b\u6539\u53d8\u4e86\u4eba\u673a\u4ea4\u4e92\u65b9\u5f0f\u3002LLMs \u7684\u8fb9\u754c\u4e5f\u4e0d\u6b62\u4e8e\u8bed\u8a00\u548c\u7b80\u5355\u95ee\u7b54\u3002\u968f\u7740\u591a\u6a21\u6001\uff08VLMs\uff09\u4e0e\u63a8\u7406\u80fd\u529b\uff08LRMs\uff09\u7684\u5174\u8d77\uff0cLLMs \u6b63\u4e0d\u65ad\u6269\u5c55\u5230\u591a\u6a21\u6001\u7406\u89e3\u3001\u751f\u6210\u4e0e\u590d\u6742\u63a8\u7406\u573a\u666f\u3002<\/p>\n\n\n\n<p>\u4f46\u6a21\u578b\u6027\u80fd\u6301\u7eed\u63d0\u5347\u7684\u80cc\u540e\uff0c\u662f\u6a21\u578b\u5c3a\u5bf8\u3001\u6570\u636e\u89c4\u6a21\u3001RL \u63a8\u7406\u957f\u5ea6\u7684\u5feb\u901f Scaling\uff0c\u662f\u7b97\u529b\u548c\u5b58\u50a8\u8d44\u6e90\u7684\u6025\u5267\u6d88\u8017\u3002\u5927\u6a21\u578b\u7684\u8bad\u7ec3\u4e0e\u63a8\u7406\u7684\u6210\u672c\u5c45\u9ad8\u4e0d\u4e0b\uff0c\u6210\u4e3a\u5236\u7ea6\u5176\u5e7f\u6cdb\u843d\u5730\u548c\u5e94\u7528\u7684\u73b0\u5b9e\u74f6\u9888\u3002<\/p>\n\n\n\n<p>\u672c\u6587\u4ece LLM \u67b6\u6784\u89d2\u5ea6\u51fa\u53d1\uff0c\u5e26\u4f60\u5256\u6790\u5927\u6a21\u578b\u7684\u6548\u7387\u79d8\u8bc0\u3002\u8fd9\u4e00\u5207\u7684\u6838\u5fc3\u5728\u4e8e Transformer \u67b6\u6784\u3002Transformer \u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u867d\u5e26\u6765\u4e86\u8fdc\u8ddd\u79bb\u5efa\u6a21\u7684\u7a81\u7834\uff0c\u5374\u56e0 O(N2) \u7684\u590d\u6742\u5ea6\u5728\u957f\u5e8f\u5217\u4efb\u52a1\u4e2d\u6210\u672c\u9ad8\u6602\u3002<strong>\u800c\u5728 RAG\u3001\u667a\u80fd\u4f53\u3001\u957f\u94fe\u63a8\u7406\u3001\u591a\u6a21\u6001\u7b49\u65b0\u5174\u573a\u666f\u4e0b\uff0c\u957f\u5e8f\u5217\u9700\u6c42\u6108\u53d1\u7a81\u51fa\uff0c\u8fdb\u4e00\u6b65\u653e\u5927\u4e86\u6548\u7387\u4e0e\u6027\u80fd\u4e4b\u95f4\u7684\u77db\u76fe\u3002<\/strong>\u540c\u65f6 Transformer \u7684 FFN \u90e8\u5206\u91c7\u7528\u5bc6\u96c6\u7684 MLP \u5c42\uff0c\u540c\u6837\u9762\u4e34\u6a21\u578b\u89c4\u6a21\u653e\u5927\u540e\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u6548\u7387\u95ee\u9898\u3002<\/p>\n\n\n\n<p>\u4f5c\u8005\uff1a\u673a\u5668\u4e4b\u5fc3<br>\u94fe\u63a5\uff1ahttps:\/\/zhuanlan.zhihu.com\/p\/1943373121248822632<br>\u8bba\u6587\u6807\u9898\uff1aSpeed Always Wins: A Survey on Efficient Architectures for Large Language Models\u8bba\u6587\u5730\u5740\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/arxiv.org\/pdf\/2508.09834\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/arxiv.org\/pdf\/2508.09834<\/a><br>\u9879\u76ee\u4ed3\u5e93\uff1a<a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/weigao266\/Awesome-Efficient-Arch\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/weigao266\/Aw<\/a><\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><em>\u56fe 2\uff1a\u5927\u8bed\u8a00\u6a21\u578b\u9ad8\u6548\u67b6\u6784\u6982\u89c8<\/em><\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"556\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/image-1024x556.png\" alt=\"\" class=\"wp-image-551\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/image-1024x556.png 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/image-300x163.png 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/image-768x417.png 768w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/image.png 1080w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u8be5\u7efc\u8ff0\u5c06\u76ee\u524d LLM \u9ad8\u6548\u67b6\u6784\u603b\u7ed3\u5206\u7c7b\u4e3a\u4ee5\u4e0b 7 \u7c7b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7ebf\u6027\u5e8f\u5217\u5efa\u6a21\uff1a\u964d\u4f4e\u6ce8\u610f\u529b\u8bad\u7ec3\u548c\u63a8\u7406\u590d\u6742\u5ea6\uff0c\u65e0\u9700 KV Cache \u5f00\u9500\u3002<\/strong><\/li>\n\n\n\n<li><strong>\u7a00\u758f\u5e8f\u5217\u5efa\u6a21\uff1a\u901a\u8fc7\u7a00\u758f\u5316\u6ce8\u610f\u529b\u77e9\u9635\uff0c\u964d\u4f4e\u8ba1\u7b97\u4e0e\u663e\u5b58\u9700\u6c42\u3002<\/strong><\/li>\n\n\n\n<li><strong>\u9ad8\u6548\u5168\u6ce8\u610f\u529b\uff1a\u5728\u4fdd\u6301\u5b8c\u6574\u6ce8\u610f\u529b\u7684\u524d\u63d0\u4e0b\u4f18\u5316\u5185\u5b58\u8bbf\u95ee\u4e0e KV \u5b58\u50a8\u3002<\/strong><\/li>\n\n\n\n<li><strong>\u7a00\u758f\u4e13\u5bb6\u6a21\u578b\uff1a\u901a\u8fc7\u6761\u4ef6\u6fc0\u6d3b\u90e8\u5206\u4e13\u5bb6\uff0c\u5927\u5e45\u63d0\u5347\u6a21\u578b\u5bb9\u91cf\u800c\u4e0d\u589e\u52a0\u7b49\u6bd4\u4f8b\u8ba1\u7b97\u6210\u672c\u3002<\/strong><\/li>\n\n\n\n<li><strong>\u6df7\u5408\u6a21\u578b\u67b6\u6784\uff1a\u7ed3\u5408\u7ebf\u6027\/\u7a00\u758f\u5e8f\u5217\u5efa\u6a21\u4e0e\u5168\u6ce8\u610f\u529b\uff0c\u517c\u987e\u6548\u7387\u4e0e\u6027\u80fd\u3002<\/strong><\/li>\n\n\n\n<li><strong>\u6269\u6563\u8bed\u8a00\u6a21\u578b\uff1a\u5229\u7528\u975e\u81ea\u56de\u5f52\u7684\u6269\u6563\u6a21\u578b\u8fdb\u884c\u8bed\u8a00\u751f\u6210\u3002<\/strong><\/li>\n\n\n\n<li><strong>\u5176\u4ed6\u6a21\u6001\u5e94\u7528\uff1a\u5c06\u8fd9\u4e9b\u9ad8\u6548\u67b6\u6784\u5e94\u7528\u4e8e\u89c6\u89c9\u3001\u8bed\u97f3\u3001\u591a\u6a21\u6001\u6a21\u578b\u3002<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"778\" height=\"1024\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/image-1-778x1024.png\" alt=\"\" class=\"wp-image-552\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/image-1-778x1024.png 778w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/image-1-228x300.png 228w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/image-1-768x1011.png 768w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/image-1.png 951w\" sizes=\"auto, (max-width: 778px) 100vw, 778px\" \/><\/figure>\n\n\n\n<p><em>\u56fe 3\uff1a\u7efc\u8ff0\u5b8c\u6574\u7ec4\u7ec7\u67b6\u6784<\/em><br><\/p>\n\n\n\n<p>\u4f5c\u8005\uff1a\u673a\u5668\u4e4b\u5fc3<br>\u94fe\u63a5\uff1ahttps:\/\/zhuanlan.zhihu.com\/p\/1943373121248822632<br>\u6765\u6e90\uff1a\u77e5\u4e4e<br>\u8457\u4f5c\u6743\u5f52\u4f5c\u8005\u6240\u6709\u3002\u5546\u4e1a\u8f6c\u8f7d\u8bf7\u8054\u7cfb\u4f5c\u8005\u83b7\u5f97\u6388\u6743\uff0c\u975e\u5546\u4e1a\u8f6c\u8f7d\u8bf7\u6ce8\u660e\u51fa\u5904\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u7ebf\u6027\u5e8f\u5217\u5efa\u6a21<\/strong><\/h3>\n\n\n\n<p><strong>\u7ebf\u6027\u5e8f\u5217\u5efa\u6a21<\/strong>\u662f\u8fd1\u5e74\u6765\u7814\u7a76\u76f8\u5f53\u706b\u70ed\u7684\u4e00\u4e2a\u65b9\u5411\uff0c\u4ee3\u8868\u6027\u5de5\u4f5c\u50cf Mamba\u3001Lighting Attention\u3001RWKV\u3001GLA\u3001TTT \u7b49\u5728\u6a21\u578b\u67b6\u6784\u65b9\u5411\u90fd\u5f15\u8d77\u8fc7\u5e7f\u6cdb\u5173\u6ce8\u3002\u6211\u4eec\u5c06\u8fd9\u7c7b\u6280\u672f\u7ec6\u5206\u4e3a\u4ee5\u4e0b\u51e0\u4e2a\u7c7b\u522b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7ebf\u6027\u6ce8\u610f\u529b<\/strong><\/li>\n\n\n\n<li><strong>\u7ebf\u6027 RNN<\/strong><\/li>\n\n\n\n<li><strong>\u72b6\u6001\u7a7a\u95f4\u6a21\u578b<\/strong><\/li>\n\n\n\n<li><strong>\u6d4b\u8bd5\u65f6\u63a8\u7406 RNN<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u5e76\u4e14\u6b63\u5982\u5728\u591a\u7bc7\u6587\u732e\u91cc\u5df2\u7ecf\u63d0\u51fa\u7684\uff0c\u8fd9\u4e9b\u7ebf\u6027\u5e8f\u5217\u5efa\u6a21\u65b9\u6cd5\u53ef\u4ee5\u6982\u62ec\u4e3a<strong>\u7edf\u4e00\u5efa\u6a21<\/strong>\u7684\u6570\u5b66\u5f62\u5f0f\uff0c\u5e76\u4e14\u80fd\u591f\u901a\u8fc7<strong>\u7ebf\u6027\u5316\u8fc7\u7a0b<\/strong>\u5c06\u9884\u8bad\u7ec3\u6a21\u578b\u6743\u91cd\u7684 Softmax Attention \u67b6\u6784\u8f6c\u4e3a Linear Sequence Modeling \u67b6\u6784\uff0c\u4ece\u800c\u83b7\u5f97\u6a21\u578b\u6548\u7387\u7684\u5927\u5e45\u63d0\u5347\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"530\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-6502e60363f83e15c33d700686b60b17_1440w.webp\" alt=\"\" class=\"wp-image-560\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-6502e60363f83e15c33d700686b60b17_1440w.webp 1080w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-6502e60363f83e15c33d700686b60b17_1440w-300x147.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-6502e60363f83e15c33d700686b60b17_1440w-1024x503.webp 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-6502e60363f83e15c33d700686b60b17_1440w-768x377.webp 768w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p><em>\u56fe 4\uff1a\u7ebf\u6027\u5e8f\u5217\u5efa\u6a21\u65b9\u6cd5<\/em><\/p>\n\n\n\n<p>\u6211\u4eec\u5c06\u5df2\u6709\u7684\u7ebf\u6027\u5e8f\u5217\u5efa\u6a21\u65b9\u6cd5\u4ece\u8bb0\u5fc6\u89c6\u89d2\u548c\u4f18\u5316\u5668\u89c6\u89d2\u5206\u522b\u8fdb\u884c\u68b3\u7406\u548c\u5bf9\u6bd4\uff0c\u8be6\u7ec6\u5f62\u5f0f\u53ef\u89c1\u4e0b\u8868\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"938\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-c276eae7c498b654a4b429492df4d635_1440w.webp\" alt=\"\" class=\"wp-image-556\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-c276eae7c498b654a4b429492df4d635_1440w.webp 1080w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-c276eae7c498b654a4b429492df4d635_1440w-300x261.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-c276eae7c498b654a4b429492df4d635_1440w-1024x889.webp 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-c276eae7c498b654a4b429492df4d635_1440w-768x667.webp 768w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p><em>\u8868 1\uff1a\u7ebf\u6027\u5e8f\u5217\u5efa\u6a21\u65b9\u6cd5\u7edf\u4e00\u5efa\u6a21\u7684 Memory \u89c6\u89d2\u548c Optimizer \u89c6\u89d2<\/em><\/p>\n\n\n\n<p>\u5176\u4e2d<strong>\u7ebf\u6027\u5316\u6280\u672f<\/strong>\u53ef\u4ee5\u8fdb\u4e00\u6b65\u7ec6\u5206\u4e3a\u57fa\u4e8e\u5fae\u8c03\u7684\u7ebf\u6027\u5316\uff0c\u548c\u57fa\u4e8e\u84b8\u998f\u7684\u7ebf\u6027\u5316\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1037\" height=\"652\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-717b3a9220496554e4fd69c8268750c7_1440w.webp\" alt=\"\" class=\"wp-image-558\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-717b3a9220496554e4fd69c8268750c7_1440w.webp 1037w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-717b3a9220496554e4fd69c8268750c7_1440w-300x189.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-717b3a9220496554e4fd69c8268750c7_1440w-1024x644.webp 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-717b3a9220496554e4fd69c8268750c7_1440w-768x483.webp 768w\" sizes=\"auto, (max-width: 1037px) 100vw, 1037px\" \/><\/figure>\n\n\n\n<p><em>\u56fe 5\uff1a\u7ebf\u6027\u5316\u65b9\u6cd5<\/em><\/p>\n\n\n\n<p>\u7efc\u8ff0\u8fd8\u8fdb\u4e00\u6b65\u603b\u7ed3\u5f52\u7eb3\u4e86\u76ee\u524d\u5728\u7ebf\u6027\u5e8f\u5217\u5efa\u6a21\u9886\u57df\u5e38\u89c1\u7684<strong>\u786c\u4ef6\u9ad8\u6548\u5b9e\u73b0\u65b9\u6cd5<\/strong>\uff0c\u53ef\u4ee5\u5f52\u7eb3\u4e3a Blelloch Scan\u3001Chunk-wise Parallel \u548c Recurrent for Inferences\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"493\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-2018c9b44f4655b1c650d8d4d4f2d7a0_1440w.webp\" alt=\"\" class=\"wp-image-557\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-2018c9b44f4655b1c650d8d4d4f2d7a0_1440w.webp 1080w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-2018c9b44f4655b1c650d8d4d4f2d7a0_1440w-300x137.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-2018c9b44f4655b1c650d8d4d4f2d7a0_1440w-1024x467.webp 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-2018c9b44f4655b1c650d8d4d4f2d7a0_1440w-768x351.webp 768w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p><em>\u56fe 6\uff1a\u7ebf\u6027\u5e8f\u5217\u5efa\u6a21\u65b9\u6cd5\u7684\u786c\u4ef6\u9ad8\u6548\u5b9e\u73b0<\/em><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u7a00\u758f\u5e8f\u5217\u5efa\u6a21<\/strong><\/h3>\n\n\n\n<p><strong>\u7a00\u758f\u5e8f\u5217\u5efa\u6a21<\/strong>\u662f\u53e6\u4e00\u7c7b\u6709\u4ee3\u8868\u6027\u7684\u9ad8\u6548\u6ce8\u610f\u529b\u673a\u5236\uff0c\u901a\u8fc7\u5229\u7528 Attention Map \u5929\u7136\u5177\u6709\u7684\u7a00\u758f\u6027\u52a0\u901f\u6ce8\u610f\u529b\u7684\u8ba1\u7b97\uff0c\u8fd9\u7c7b\u65b9\u6cd5\u53ef\u4ee5\u8fdb\u4e00\u6b65\u7ec6\u5206\u4e3a\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u9759\u6001\u7a00\u758f\u6ce8\u610f\u529b<\/strong><\/li>\n\n\n\n<li><strong>\u52a8\u6001\u7a00\u758f\u6ce8\u610f\u529b<\/strong><\/li>\n\n\n\n<li><strong>\u514d\u8bad\u7ec3\u7a00\u758f\u6ce8\u610f\u529b<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u4ee3\u8868\u6027\u7684\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6cd5\u5982 Global Attention\u3001Window Attention\u3001Dilated Attention \u7b49\uff0c\u53ca\u5176\u5de5\u4f5c\u539f\u7406\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"365\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-6b609513d5522bb892fcf8289084ce54_1440w.webp\" alt=\"\" class=\"wp-image-563\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-6b609513d5522bb892fcf8289084ce54_1440w.webp 1080w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-6b609513d5522bb892fcf8289084ce54_1440w-300x101.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-6b609513d5522bb892fcf8289084ce54_1440w-1024x346.webp 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-6b609513d5522bb892fcf8289084ce54_1440w-768x260.webp 768w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p><em>\u56fe 7\uff1a\u7a00\u758f\u6ce8\u610f\u529b\u7684\u51e0\u79cd\u7ecf\u5178\u5f62\u5f0f<\/em><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u9ad8\u6548\u5168\u6ce8\u610f\u529b<\/strong><\/h3>\n\n\n\n<p>\u53e6\u4e00\u7c7b\u9ad8\u6548\u6ce8\u610f\u529b\u7b97\u6cd5\u53ef\u4ee5\u7edf\u4e00\u5f52\u7eb3\u4e3a<strong>\u9ad8\u6548\u5168\u6ce8\u610f\u529b<\/strong>\uff0c\u8fd9\u7c7b\u65b9\u6cd5\u53ef\u4ee5\u6839\u636e\u7b97\u6cd5\u601d\u8def\u8fdb\u4e00\u6b65\u7ec6\u5206\u4e3a\u5982\u4e0b\u51e0\u7c7b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>IO-Aware Attention<\/strong><\/li>\n\n\n\n<li><strong>Grouped Attention<\/strong><\/li>\n\n\n\n<li><strong>Mixture of Attention<\/strong><\/li>\n\n\n\n<li><strong>Quantized Attention<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u5176\u4e2d IO-Aware Attention \u6307\u4ee3\u76ee\u524d\u4f7f\u7528\u975e\u5e38\u5e7f\u6cdb\u7684 Flash Attention \u7cfb\u5217\u5de5\u4f5c\uff0cGrouped Attention \u5219\u5305\u542b\u5e7f\u4e3a\u4f7f\u7528\u7684 GQA\u3001MLA \u7b49\u5168\u6ce8\u610f\u529b\u53d8\u4f53\uff0c\u51e0\u79cd\u4ee3\u8868\u6027\u65b9\u6cd5\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"830\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-b468531fe46a2d28f27e2adc8ff77187_1440w.webp\" alt=\"\" class=\"wp-image-562\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-b468531fe46a2d28f27e2adc8ff77187_1440w.webp 1080w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-b468531fe46a2d28f27e2adc8ff77187_1440w-300x231.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-b468531fe46a2d28f27e2adc8ff77187_1440w-1024x787.webp 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-b468531fe46a2d28f27e2adc8ff77187_1440w-768x590.webp 768w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p><em>\u56fe 8\uff1aGrouped Attention \u7684\u51e0\u79cd\u4ee3\u8868\u6027\u65b9\u6cd5<\/em><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u7a00\u758f\u6df7\u5408\u4e13\u5bb6<\/strong><\/h3>\n\n\n\n<p><strong>\u7a00\u758f\u6df7\u5408\u4e13\u5bb6<\/strong>\u662f\u5bf9 Transformer \u67b6\u6784\u4e2d\u53e6\u4e00\u4e2a\u91cd\u8981\u6a21\u5757 FFN \u505a\u7684\u4e00\u7c7b\u91cd\u8981\u6539\u8fdb\uff0c\u5df2\u7ecf\u9010\u6e10\u6210\u4e3a\uff08\u8bed\u8a00\u548c\u591a\u6a21\u6001\uff09\u5927\u6a21\u578b\u67b6\u6784\u7684\u4e8b\u5b9e\u6807\u51c6\u3002\u7efc\u8ff0\u4e2d\u5c06\u76f8\u5173\u6587\u732e\u6309\u4ee5\u4e0b\u4e09\u4e2a\u65b9\u5411\u8fdb\u884c\u5206\u7c7b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Routing Mechanisms<\/strong><\/li>\n\n\n\n<li><strong>Expert Architectures<\/strong><\/li>\n\n\n\n<li><strong>MoE Conversion<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u8def\u7531\u673a\u5236\u5305\u62ec Token-choice \u548c Expert-choice \u4e24\u7c7b\uff0c\u5176\u539f\u7406\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1005\" height=\"350\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-f5708c175f202afcbe951119ba3bcca1_1440w.webp\" alt=\"\" class=\"wp-image-559\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-f5708c175f202afcbe951119ba3bcca1_1440w.webp 1005w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-f5708c175f202afcbe951119ba3bcca1_1440w-300x104.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-f5708c175f202afcbe951119ba3bcca1_1440w-768x267.webp 768w\" sizes=\"auto, (max-width: 1005px) 100vw, 1005px\" \/><\/figure>\n\n\n\n<p><em>\u56fe 9\uff1aMoE \u8def\u7531\u673a\u5236<\/em><\/p>\n\n\n\n<p>\u4e13\u5bb6\u7ed3\u6784\u7684\u521b\u65b0\u5de5\u4f5c\u5305\u62ec\uff1a\u5171\u4eab\u4e13\u5bb6\u3001\u7ec6\u7c92\u5ea6\u4e13\u5bb6\u3001\u96f6\u4e13\u5bb6\u3001\u6df1\u5ea6\u8def\u7531\u7b49\uff0c\u5176\u4f5c\u7528\u548c\u539f\u7406\u53ef\u89c1\u4e0b\u56fe\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"973\" height=\"537\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-d27eb1f1453803c457e6aea16d957856_1440w.webp\" alt=\"\" class=\"wp-image-553\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-d27eb1f1453803c457e6aea16d957856_1440w.webp 973w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-d27eb1f1453803c457e6aea16d957856_1440w-300x166.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-d27eb1f1453803c457e6aea16d957856_1440w-768x424.webp 768w\" sizes=\"auto, (max-width: 973px) 100vw, 973px\" \/><\/figure>\n\n\n\n<p><em>\u56fe 10\uff1aMoE \u4e13\u5bb6\u67b6\u6784<\/em><\/p>\n\n\n\n<p>\u53e6\u5916\u4e00\u4e2a\u91cd\u8981\u7684\u65b9\u5411\u662f MoE \u8f6c\u6362\uff0c\u5df2\u6709\u7684\u5de5\u4f5c\u5305\u62ec\u901a\u8fc7 Split\u3001Copy\u3001Merge \u7b49\u624b\u6bb5\u5bf9\u4e13\u5bb6\u8fdb\u884c\u6784\u9020\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"217\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-70a7a466c5357ea0708b2b98d1bdc6a4_1440w.webp\" alt=\"\" class=\"wp-image-564\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-70a7a466c5357ea0708b2b98d1bdc6a4_1440w.webp 1080w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-70a7a466c5357ea0708b2b98d1bdc6a4_1440w-300x60.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-70a7a466c5357ea0708b2b98d1bdc6a4_1440w-1024x206.webp 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-70a7a466c5357ea0708b2b98d1bdc6a4_1440w-768x154.webp 768w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p><em>\u56fe 11\uff1aMoE \u8f6c\u5316\u673a\u5236<\/em><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u6df7\u5408\u67b6\u6784<\/strong><\/h3>\n\n\n\n<p><strong>\u6df7\u5408\u67b6\u6784<\/strong>\u662f\u8fd1\u5e74\u6765\u51fa\u73b0\u7684\u4e00\u79cd\u5b9e\u7528\u7684\u65b0\u578b\u67b6\u6784\uff0c\u53ef\u4ee5\u5728\u7ebf\u6027\/\u7a00\u758f\u6ce8\u610f\u529b\u548c\u5168\u6ce8\u610f\u529b\u4e4b\u95f4\u53d6\u5f97\u5fae\u5999\u7684 Trade-off\uff0c\u4e5f\u5728\u6548\u7387\u548c\u6548\u679c\u95f4\u627e\u5230\u4e86\u6700\u4f73\u751c\u871c\u70b9\u3002\u5177\u4f53\u53ef\u7ec6\u5206\u4e3a\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5c42\u95f4\u6df7\u5408<\/strong><\/li>\n\n\n\n<li><strong>\u5c42\u5185\u6df7\u5408<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"874\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-c464c76f890ce22f4cab086e7ff988fb_1440w.webp\" alt=\"\" class=\"wp-image-561\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-c464c76f890ce22f4cab086e7ff988fb_1440w.webp 1080w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-c464c76f890ce22f4cab086e7ff988fb_1440w-300x243.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-c464c76f890ce22f4cab086e7ff988fb_1440w-1024x829.webp 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-c464c76f890ce22f4cab086e7ff988fb_1440w-768x622.webp 768w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n\n\n\n<p><em>\u56fe 12\uff1a\u6df7\u5408\u67b6\u6784\u5f62\u5f0f<\/em><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u6269\u6563\u5927\u8bed\u8a00\u6a21\u578b<\/strong><\/h3>\n\n\n\n<p><strong>\u6269\u6563\u5927\u8bed\u8a00\u6a21\u578b<\/strong>\u662f\u8fd1\u671f\u51fa\u73b0\u7684\u4e00\u4e2a\u70ed\u95e8\u65b9\u5411\uff0c\u521b\u65b0\u6027\u5730\u5c06\u6269\u6563\u6a21\u578b\u4ece\u89c6\u89c9\u751f\u6210\u4efb\u52a1\u8fc1\u79fb\u81f3\u8bed\u8a00\u4efb\u52a1\uff0c\u4ece\u800c\u5728\u8bed\u8a00\u751f\u6210\u901f\u5ea6\u4e0a\u53d6\u5f97\u5927\u5e45\u8fdb\u6b65\u3002\u76f8\u5173\u5de5\u4f5c\u53ef\u4ee5\u7ec6\u5206\u4e3a\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Non-Autoregressive Diffusion LLM<\/strong><\/li>\n\n\n\n<li><strong>Bridging Diffusion LLM and Autoregressive<\/strong><\/li>\n\n\n\n<li><strong>Extending Diffusion LLM to Multimodality<\/strong><\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"946\" height=\"431\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-a61b57987fb19507955eb1f421273c13_1440w.webp\" alt=\"\" class=\"wp-image-554\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-a61b57987fb19507955eb1f421273c13_1440w.webp 946w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-a61b57987fb19507955eb1f421273c13_1440w-300x137.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-a61b57987fb19507955eb1f421273c13_1440w-768x350.webp 768w\" sizes=\"auto, (max-width: 946px) 100vw, 946px\" \/><\/figure>\n\n\n\n<p><em>\u56fe 13\uff1a\u6269\u6563\u5927\u8bed\u8a00\u6a21\u578b\u673a\u5236<\/em><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u5e94\u7528\u81f3\u5176\u4ed6\u6a21\u6001<\/strong><\/h3>\n\n\n\n<p>\u6700\u540e\u4e00\u4e2a\u91cd\u8981\u7684\u90e8\u5206\u662f<strong>\u9ad8\u6548\u67b6\u6784\u5728\u5176\u4ed6\u6a21\u6001\u4e0a\u7684\u5e94\u7528\uff0c\u6db5\u76d6\u89c6\u89c9\u3001\u97f3\u9891\u548c\u591a\u6a21\u6001<\/strong>\u3002\u4ee5 Mamba \u4e3a\u4ee3\u8868\u7684\u7ebf\u6027\u6a21\u578b\u88ab\u5e7f\u6cdb\u5e94\u7528\u81f3\u591a\u79cd\u6a21\u6001\u4efb\u52a1\u4e0a\uff0c\u5e76\u53d6\u5f97\u4e86\u4f18\u79c0\u7684\u8868\u73b0\uff0c\u7efc\u8ff0\u5c06\u8fd9\u7c7b\u6a21\u578b\u603b\u7ed3\u68b3\u7406\u81f3\u5982\u4e0b\u8868\u683c\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"688\" src=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-490e6a8881ca52cd34f059c8df27bdfc_1440w.webp\" alt=\"\" class=\"wp-image-555\" srcset=\"https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-490e6a8881ca52cd34f059c8df27bdfc_1440w.webp 1080w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-490e6a8881ca52cd34f059c8df27bdfc_1440w-300x191.webp 300w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-490e6a8881ca52cd34f059c8df27bdfc_1440w-1024x652.webp 1024w, https:\/\/aitimes.link\/wp-content\/uploads\/2025\/08\/v2-490e6a8881ca52cd34f059c8df27bdfc_1440w-768x489.webp 768w\" sizes=\"auto, (max-width: 1080px) 100vw, 1080px\" \/><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>\u8fd1\u5e74\u6765\uff0c\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u8bed\u8a00\u7406\u89e3\u4e0e\u751f\u6210\u80fd\u529b\uff0c\u63a8\u52a8\u4e86\u6587\u672c\u751f\u6210\u3001\u4ee3\u7801\u751f\u6210\u3001\u95ee\u7b54\u3001\u7ffb\u8bd1\u7b49\u4efb\u52a1\u7684\u7a81\u7834 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[],"class_list":["post-550","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts\/550","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/comments?post=550"}],"version-history":[{"count":2,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts\/550\/revisions"}],"predecessor-version":[{"id":566,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/posts\/550\/revisions\/566"}],"wp:attachment":[{"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/media?parent=550"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/categories?post=550"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aitimes.link\/index.php\/wp-json\/wp\/v2\/tags?post=550"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}