{"id":1072581,"date":"2023-07-18T22:35:00","date_gmt":"2023-07-19T05:35:00","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&#038;p=1072581"},"modified":"2024-09-25T03:54:38","modified_gmt":"2024-09-25T10:54:38","slug":"retnet","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/retnet\/","title":{"rendered":"\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa\u5168\u65b0\u5927\u6a21\u578b\u57fa\u7840\u67b6\u6784RetNet\uff0c\u6216\u5c06\u6210\u4e3aTransformer\u6709\u529b\u7ee7\u627f\u8005\uff01"},"content":{"rendered":"\n<p>\u4f5c\u8005\uff1a\u91cf\u5b50\u4f4d<\/p>\n\n\n\n<p><em>\u4f5c\u4e3a\u5168\u65b0\u7684\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\uff0cRetNet \u540c\u65f6\u5b9e\u73b0\u4e86\u826f\u597d\u7684\u6269\u5c55\u7ed3\u679c\u3001\u5e76\u884c\u8bad\u7ec3\u3001\u4f4e\u6210\u672c\u90e8\u7f72\u548c\u9ad8\u6548\u63a8\u7406\u3002\u8fd9\u4e9b\u7279\u6027\u5c06\u4f7f RetNet \u6709\u53ef\u80fd\u6210\u4e3a\u7ee7 Transformer \u4e4b\u540e\u5927\u8bed\u8a00\u6a21\u578b\u57fa\u7840\u7f51\u7edc\u67b6\u6784\u7684\u6709\u529b\u7ee7\u627f\u8005\u3002<\/em><\/p>\n\n\n\n<p><em>\u2014\u2014\u97e6\u798f\u5982\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u5168\u7403\u7814\u7a76\u5408\u4f19\u4eba<\/em><\/p>\n\n\n\n<p>\u4ee5\u4e0b\u5185\u5bb9\u7ecf\u6388\u6743\u8f6c\u8f7d\u81ea\u516c\u4f17\u53f7\u201c\u91cf\u5b50\u4f4d\u201d\uff0c\u539f\u6587\u6807\u9898\u300aTransformer\u540e\u7ee7\u6709\u6a21\uff01MSRA\u63d0\u51fa\u5168\u65b0\u5927\u6a21\u578b\u57fa\u7840\u67b6\u6784\uff1a\u63a8\u7406\u901f\u5ea68\u500d\u63d0\u5347\uff0c\u5185\u5b58\u5360\u7528\u51cf\u5c1170%\u300b\u3002<\/p>\n\n\n\n<p>\u4eca\u5929\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e0a\u7ebf\u4e86\u5927\u6a21\u578b\u65b0\u67b6\u6784\u7684\u8bba\u6587\u201cRetentive Network: A Successor to Transformer for Large Language Models\u201d\uff0c\u8be5\u57fa\u7840\u67b6\u6784\u91c7\u7528\u4e86\u65b0\u7684 Retention \u673a\u5236\u6765\u4ee3\u66ff Attention\uff0c\u5411 Transformer \u53d1\u8d77\u6311\u6218\uff01<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"953\" height=\"310\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-1.png\" alt=\"paper screenshot\" class=\"wp-image-1072782\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-1.png 953w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-1-300x98.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-1-768x250.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-1-240x78.png 240w\" sizes=\"auto, (max-width: 953px) 100vw, 953px\" \/><\/figure>\n\n\n\n<p>\u5b9e\u9a8c\u6570\u636e\u4e5f\u663e\u793a\uff0c\u5728\u8bed\u8a00\u5efa\u6a21\u4efb\u52a1\u4e0a\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>RetNet \u53ef\u4ee5\u8fbe\u5230\u4e0e Transformer \u76f8\u5f53\u7684\u56f0\u60d1\u5ea6\uff08perplexity\uff09<\/li>\n\n\n\n<li>\u63a8\u7406\u901f\u5ea6\u8fbe8.4\u500d<\/li>\n\n\n\n<li>\u5185\u5b58\u5360\u7528\u51cf\u5c1170%<\/li>\n\n\n\n<li>\u5177\u6709\u826f\u597d\u7684\u6269\u5c55\u6027<\/li>\n<\/ul>\n\n\n\n<p>\u5e76\u4e14\u5f53\u6a21\u578b\u5927\u5c0f\u5927\u4e8e\u4e00\u5b9a\u89c4\u6a21\u65f6\uff0cRetNet \u7684\u8868\u73b0\u4f1a\u4f18\u4e8e Transformer\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"960\" height=\"425\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-2.png\" alt=\"table\" class=\"wp-image-1072785\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-2.png 960w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-2-300x133.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-2-768x340.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-2-240x106.png 240w\" sizes=\"auto, (max-width: 960px) 100vw, 960px\" \/><\/figure>\n\n\n\n<p>\u5177\u4f53\u8be6\u60c5\uff0c\u4e00\u8d77\u6765\u770b\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u89e3\u51b3-\u4e0d\u53ef\u80fd\u4e09\u89d2\">\u89e3\u51b3\u201c\u4e0d\u53ef\u80fd\u4e09\u89d2\u201d<\/h2>\n\n\n\n<p>Transformer \u5728\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u91cd\u8981\u6027\u6bcb\u5eb8\u7f6e\u7591\u3002\u65e0\u8bba\u662f OpenAI \u7684 GPT \u7cfb\u5217\uff0c\u8fd8\u662f\u8c37\u6b4c\u7684 PaLM\u3001Meta \u7684 LLaMA\uff0c\u90fd\u662f\u57fa\u4e8e Transformer \u6253\u9020\u3002<\/p>\n\n\n\n<p>\u4f46 Transformer \u4e5f\u5e76\u975e\u5b8c\u7f8e\u65e0\u7f3a\uff1a\u5176\u5e76\u884c\u5904\u7406\u673a\u5236\u662f\u4ee5\u4f4e\u6548\u63a8\u7406\u4e3a\u4ee3\u4ef7\u7684\uff0c\u6bcf\u4e2a\u6b65\u9aa4\u7684\u590d\u6742\u5ea6\u4e3a O(N)\uff1bTransformer \u662f\u5185\u5b58\u5bc6\u96c6\u578b\u6a21\u578b\uff0c\u5e8f\u5217\u8d8a\u957f\uff0c\u5360\u7528\u7684\u5185\u5b58\u8d8a\u591a\u3002<\/p>\n\n\n\n<p>\u5728\u6b64\u4e4b\u524d\uff0c\u5927\u5bb6\u4e5f\u4e0d\u662f\u6ca1\u60f3\u8fc7\u7ee7\u7eed\u6539\u8fdb Transformer\u3002\u4f46\u4e3b\u8981\u7684\u51e0\u79cd\u7814\u7a76\u65b9\u5411\u90fd\u6709\u4e9b\u987e\u6b64\u5931\u5f7c\uff1a<\/p>\n\n\n\n<p>\u7ebf\u6027 Attention \u53ef\u4ee5\u964d\u4f4e\u63a8\u7406\u6210\u672c\uff0c\u4f46\u6027\u80fd\u8f83\u5dee\uff1b<\/p>\n\n\n\n<p>\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u5219\u65e0\u6cd5\u8fdb\u884c\u5e76\u884c\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p>\u4e5f\u5c31\u662f\u8bf4\uff0c\u8fd9\u4e9b\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\u9762\u524d\u6446\u7740\u4e00\u4e2a\u201c\u4e0d\u53ef\u80fd\u4e09\u89d2\u201d\uff0c\u4e09\u4e2a\u89d2\u4ee3\u8868\u7684\u5206\u522b\u662f\uff1a\u5e76\u884c\u8bad\u7ec3\u3001\u4f4e\u6210\u672c\u63a8\u7406\u548c\u826f\u597d\u7684\u6269\u5c55\u6027\u80fd\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"868\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-3-1024x868.png\" alt=\"diagram\" class=\"wp-image-1072788\" style=\"width:614px;height:auto\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-3-1024x868.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-3-300x254.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-3-768x651.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-3-212x180.png 212w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-3.png 1066w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>RetNet \u7684\u7814\u7a76\u4eba\u5458\u60f3\u505a\u7684\uff0c\u5c31\u662f\u5316\u4e0d\u53ef\u80fd\u4e3a\u53ef\u80fd\u3002<\/p>\n\n\n\n<p>\u5177\u4f53\u800c\u8a00\uff0cRetNet \u5728 Transformer \u7684\u57fa\u7840\u4e0a\uff0c\u4f7f\u7528\u591a\u5c3a\u5ea6\u4fdd\u6301\uff08Retention\uff09\u673a\u5236\u66ff\u4ee3\u4e86\u6807\u51c6\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u3002<\/p>\n\n\n\n<p>\u4e0e\u6807\u51c6\u81ea\u6ce8\u610f\u529b\u673a\u5236\u76f8\u6bd4\uff0c\u4fdd\u6301\u673a\u5236\u6709\u51e0\u5927\u7279\u70b9\uff1a<\/p>\n\n\n\n<p>\u5f15\u5165\u4f4d\u7f6e\u76f8\u5173\u7684\u6307\u6570\u8870\u51cf\u9879\u53d6\u4ee3 softmax\uff0c\u7b80\u5316\u4e86\u8ba1\u7b97\uff0c\u540c\u65f6\u4f7f\u524d\u6b65\u7684\u4fe1\u606f\u4ee5\u8870\u51cf\u7684\u5f62\u5f0f\u4fdd\u7559\u4e0b\u6765\u3002<\/p>\n\n\n\n<p>\u5f15\u5165\u590d\u6570\u7a7a\u95f4\u8868\u8fbe\u4f4d\u7f6e\u4fe1\u606f\uff0c\u53d6\u4ee3\u7edd\u5bf9\u6216\u76f8\u5bf9\u4f4d\u7f6e\u7f16\u7801\uff0c\u5bb9\u6613\u8f6c\u6362\u4e3a\u9012\u5f52\u5f62\u5f0f\u3002<\/p>\n\n\n\n<p>\u53e6\u5916\uff0c\u4fdd\u6301\u673a\u5236\u4f7f\u7528\u591a\u5c3a\u5ea6\u7684\u8870\u51cf\u7387\uff0c\u589e\u52a0\u4e86\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\uff0c\u5e76\u5229\u7528 GroupNorm \u7684\u7f29\u653e\u4e0d\u53d8\u6027\u6765\u63d0\u9ad8 Retention \u5c42\u7684\u6570\u503c\u7cbe\u5ea6\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"952\" height=\"445\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-4.png\" alt=\"diagram\" class=\"wp-image-1072791\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-4.png 952w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-4-300x140.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-4-768x359.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-4-240x112.png 240w\" sizes=\"auto, (max-width: 952px) 100vw, 952px\" \/><\/figure>\n\n\n\n<p id=\"caption-attachment-45554\">RetNet \u7684\u53cc\u91cd\u8868\u793a<\/p>\n\n\n\n<p>\u6bcf\u4e2a RetNet \u5757\u5305\u542b\u4e24\u4e2a\u6a21\u5757\uff1a\u591a\u5c3a\u5ea6\u4fdd\u6301\uff08MSR\uff09\u6a21\u5757\u548c\u524d\u9988\u7f51\u7edc\uff08FFN\uff09\u6a21\u5757\u3002<\/p>\n\n\n\n<p>\u4fdd\u6301\u673a\u5236\u652f\u6301\u4ee5\u4e09\u79cd\u5f62\u5f0f\u8868\u793a\u5e8f\u5217\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5e76\u884c<\/li>\n\n\n\n<li>\u9012\u5f52<\/li>\n\n\n\n<li>\u5206\u5757\u9012\u5f52\uff0c\u5373\u5e76\u884c\u8868\u793a\u548c\u9012\u5f52\u8868\u793a\u7684\u6df7\u5408\u5f62\u5f0f\uff0c\u5c06\u8f93\u5165\u5e8f\u5217\u5212\u5206\u4e3a\u5757\uff0c\u5728\u5757\u5185\u6309\u7167\u5e76\u884c\u8868\u793a\u8fdb\u884c\u8ba1\u7b97\uff0c\u5728\u5757\u95f4\u9075\u5faa\u9012\u5f52\u8868\u793a\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5176\u4e2d\uff0c\u5e76\u884c\u8868\u793a\u4f7f RetNet \u53ef\u4ee5\u50cf Transformer \u4e00\u6837\u9ad8\u6548\u5730\u5229\u7528 GPU \u8fdb\u884c\u5e76\u884c\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p>\u9012\u5f52\u8868\u793a\u5b9e\u73b0\u4e86O(1)\u7684\u63a8\u7406\u590d\u6742\u5ea6\uff0c\u964d\u4f4e\u4e86\u5185\u5b58\u5360\u7528\u548c\u5ef6\u8fdf\u3002<\/p>\n\n\n\n<p>\u5206\u5757\u9012\u5f52\u5219\u53ef\u4ee5\u66f4\u9ad8\u6548\u5730\u5904\u7406\u957f\u5e8f\u5217\u3002<\/p>\n\n\n\n<p>\u8fd9\u6837\u4e00\u6765\uff0cRetNet \u5c31\u4f7f\u5f97\u201c\u4e0d\u53ef\u80fd\u4e09\u89d2\u201d\u6210\u4e3a\u53ef\u80fd\u3002\u4ee5\u4e0b\u4e3a RetNet \u4e0e\u5176\u4ed6\u57fa\u7840\u67b6\u6784\u7684\u5bf9\u6bd4\u7ed3\u679c\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"358\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-5-1024x358.png\" alt=\"table\" class=\"wp-image-1072794\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-5-1024x358.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-5-300x105.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-5-768x268.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-5-240x84.png 240w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-5.png 1073w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u5728\u8bed\u8a00\u5efa\u6a21\u4efb\u52a1\u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c\uff0c\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86 RetNet \u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<p>\u7ed3\u679c\u663e\u793a\uff0cRetNet \u53ef\u4ee5\u8fbe\u5230\u4e0e Transformer \u76f8\u4f3c\u7684\u56f0\u60d1\u5ea6\uff08PPL\uff0c\u8bc4\u4ef7\u8bed\u8a00\u6a21\u578b\u597d\u574f\u7684\u6307\u6807\uff0c\u8d8a\u5c0f\u8d8a\u597d\uff09\u3002<\/p>\n\n\n\n<p>\u540c\u65f6\uff0c\u5728\u6a21\u578b\u53c2\u6570\u4e3a70\u4ebf\u3001\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u4e3a8k\u7684\u60c5\u51b5\u4e0b\uff0cRetNet \u7684\u63a8\u7406\u901f\u5ea6\u80fd\u8fbe\u5230 Transformer \u76848.4\u500d\uff0c\u5185\u5b58\u5360\u7528\u51cf\u5c1170%\u3002<\/p>\n\n\n\n<p>\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0cRetNet \u5728\u5185\u5b58\u8282\u7701\u548c\u52a0\u901f\u6548\u679c\u65b9\u9762\uff0c\u4e5f\u6bd4\u6807\u51c6 Transformer+FlashAttention \u8868\u73b0\u66f4\u597d\uff0c\u5206\u522b\u8fbe\u523025-50%\u548c7\u500d\u3002<\/p>\n\n\n\n<p>\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cRetNet \u7684\u63a8\u7406\u6210\u672c\u4e0e\u5e8f\u5217\u957f\u5ea6\u65e0\u5173\uff0c\u63a8\u7406\u5ef6\u8fdf\u5bf9\u6279\u91cf\u5927\u5c0f\u4e0d\u654f\u611f\uff0c\u5141\u8bb8\u9ad8\u541e\u5410\u91cf\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1020\" height=\"808\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-6.png\" alt=\"diagram\" class=\"wp-image-1072797\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-6.png 1020w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-6-300x238.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-6-768x608.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-6-227x180.png 227w\" sizes=\"auto, (max-width: 1020px) 100vw, 1020px\" \/><\/figure>\n\n\n\n<p>\u53e6\u5916\uff0c\u5f53\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u5927\u4e8e20\u4ebf\u65f6\uff0cRetNet \u7684\u8868\u73b0\u4f1a\u4f18\u4e8e Transformer\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"386\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-7-1024x386.png\" alt=\"chart, line chart\" class=\"wp-image-1072800\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-7-1024x386.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-7-300x113.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-7-768x289.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-7-240x90.png 240w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2024\/08\/retnet-7.png 1027w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>Retentive Network: A Successor to Transformer for Large Language Models<\/p>\n\n\n\n<p>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2307.08621<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f5c\u8005\uff1a\u91cf\u5b50\u4f4d \u4f5c\u4e3a\u5168\u65b0\u7684\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\uff0cRetNet \u540c\u65f6\u5b9e\u73b0\u4e86\u826f\u597d\u7684\u6269\u5c55\u7ed3\u679c\u3001\u5e76\u884c\u8bad\u7ec3\u3001\u4f4e\u6210\u672c\u90e8\u7f72\u548c\u9ad8\u6548\u63a8\u7406\u3002\u8fd9\u4e9b\u7279\u6027\u5c06\u4f7f RetNet \u6709\u53ef\u80fd\u6210\u4e3a\u7ee7 Transformer \u4e4b\u540e\u5927\u8bed\u8a00\u6a21\u578b\u57fa\u7840\u7f51\u7edc\u67b6\u6784\u7684\u6709\u529b\u7ee7\u627f\u8005\u3002 \u2014\u2014\u97e6\u798f\u5982\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u5168\u7403\u7814\u7a76\u5408\u4f19\u4eba \u4ee5\u4e0b\u5185\u5bb9\u7ecf\u6388\u6743\u8f6c\u8f7d\u81ea\u516c\u4f17\u53f7\u201c\u91cf\u5b50\u4f4d\u201d\uff0c\u539f\u6587\u6807\u9898\u300aTransformer\u540e\u7ee7\u6709\u6a21\uff01MSRA\u63d0\u51fa\u5168\u65b0\u5927\u6a21\u578b\u57fa\u7840\u67b6\u6784\uff1a\u63a8\u7406\u901f\u5ea68\u500d\u63d0\u5347\uff0c\u5185\u5b58\u5360\u7528\u51cf\u5c1170%\u300b\u3002 \u4eca\u5929\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u4e0a\u7ebf\u4e86\u5927\u6a21\u578b\u65b0\u67b6\u6784\u7684\u8bba\u6587\u201cRetentive Network: A Successor to Transformer for Large Language Models\u201d\uff0c\u8be5\u57fa\u7840\u67b6\u6784\u91c7\u7528\u4e86\u65b0\u7684 Retention \u673a\u5236\u6765\u4ee3\u66ff Attention\uff0c\u5411 Transformer \u53d1\u8d77\u6311\u6218\uff01 \u5b9e\u9a8c\u6570\u636e\u4e5f\u663e\u793a\uff0c\u5728\u8bed\u8a00\u5efa\u6a21\u4efb\u52a1\u4e0a\uff1a \u5e76\u4e14\u5f53\u6a21\u578b\u5927\u5c0f\u5927\u4e8e\u4e00\u5b9a\u89c4\u6a21\u65f6\uff0cRetNet \u7684\u8868\u73b0\u4f1a\u4f18\u4e8e Transformer\u3002 \u5177\u4f53\u8be6\u60c5\uff0c\u4e00\u8d77\u6765\u770b\u3002 Transformer \u5728\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u91cd\u8981\u6027\u6bcb\u5eb8\u7f6e\u7591\u3002\u65e0\u8bba\u662f OpenAI \u7684 GPT \u7cfb\u5217\uff0c\u8fd8\u662f\u8c37\u6b4c\u7684 PaLM\u3001Meta \u7684 LLaMA\uff0c\u90fd\u662f\u57fa\u4e8e Transformer \u6253\u9020\u3002 \u4f46 Transformer \u4e5f\u5e76\u975e\u5b8c\u7f8e\u65e0\u7f3a\uff1a\u5176\u5e76\u884c\u5904\u7406\u673a\u5236\u662f\u4ee5\u4f4e\u6548\u63a8\u7406\u4e3a\u4ee3\u4ef7\u7684\uff0c\u6bcf\u4e2a\u6b65\u9aa4\u7684\u590d\u6742\u5ea6\u4e3a O(N)\uff1bTransformer \u662f\u5185\u5b58\u5bc6\u96c6\u578b\u6a21\u578b\uff0c\u5e8f\u5217\u8d8a\u957f\uff0c\u5360\u7528\u7684\u5185\u5b58\u8d8a\u591a\u3002 \u5728\u6b64\u4e4b\u524d\uff0c\u5927\u5bb6\u4e5f\u4e0d\u662f\u6ca1\u60f3\u8fc7\u7ee7\u7eed\u6539\u8fdb Transformer\u3002\u4f46\u4e3b\u8981\u7684\u51e0\u79cd\u7814\u7a76\u65b9\u5411\u90fd\u6709\u4e9b\u987e\u6b64\u5931\u5f7c\uff1a \u7ebf\u6027 Attention \u53ef\u4ee5\u964d\u4f4e\u63a8\u7406\u6210\u672c\uff0c\u4f46\u6027\u80fd\u8f83\u5dee\uff1b \u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u5219\u65e0\u6cd5\u8fdb\u884c\u5e76\u884c\u8bad\u7ec3\u3002 \u4e5f\u5c31\u662f\u8bf4\uff0c\u8fd9\u4e9b\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\u9762\u524d\u6446\u7740\u4e00\u4e2a\u201c\u4e0d\u53ef\u80fd\u4e09\u89d2\u201d\uff0c\u4e09\u4e2a\u89d2\u4ee3\u8868\u7684\u5206\u522b\u662f\uff1a\u5e76\u884c\u8bad\u7ec3\u3001\u4f4e\u6210\u672c\u63a8\u7406\u548c\u826f\u597d\u7684\u6269\u5c55\u6027\u80fd\u3002 RetNet \u7684\u7814\u7a76\u4eba\u5458\u60f3\u505a\u7684\uff0c\u5c31\u662f\u5316\u4e0d\u53ef\u80fd\u4e3a\u53ef\u80fd\u3002 \u5177\u4f53\u800c\u8a00\uff0cRetNet [&hellip;]<\/p>\n","protected":false},"author":34512,"featured_media":1072779,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1072581","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072581","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":3,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072581\/revisions"}],"predecessor-version":[{"id":1072803,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072581\/revisions\/1072803"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1072779"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1072581"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1072581"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1072581"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1072581"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}