{"id":4914,"date":"2024-10-20T16:30:30","date_gmt":"2024-10-20T08:30:30","guid":{"rendered":"https:\/\/nullthought.net\/?p=4914"},"modified":"2024-11-07T11:00:44","modified_gmt":"2024-11-07T03:00:44","slug":"%e5%88%a9%e7%94%a8%e9%97%aa%e5%ad%98%ef%bc%88flash%ef%bc%89%e5%ad%98%e5%82%a8%e6%a8%a1%e5%9e%8b%e5%8f%82%e6%95%b0%e5%b9%b6%e6%8c%89%e9%9c%80%e5%8a%a0%e8%bd%bd%ef%bc%8c%e5%9c%a8%e6%9c%89%e9%99%90","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=4914","title":{"rendered":"\u5229\u7528\u95ea\u5b58\uff08flash\uff09\u5b58\u50a8\u6a21\u578b\u53c2\u6570\u5e76\u6309\u9700\u52a0\u8f7d\uff0c\u5728\u6709\u9650\u5185\u5b58\u5185\u8fd0\u884c\u5927\u6a21\u578b\uff08LLMs\uff09"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2312.11514\" target=\"_blank\" rel=\"noreferrer noopener\">LLM in a flash: Efficient Large Language Model Inference with Limited Memory<\/a><\/strong>\u63a2\u8ba8\u4e86\u5728\u5185\u5b58\u53d7\u9650\u7684\u786c\u4ef6\u8bbe\u5907\u4e0a\u5982\u4f55\u9ad8\u6548\u8fd0\u884c\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08Large Language Models, LLMs\uff09\uff0c\u5c24\u5176\u662f\u5229\u7528\u95ea\u5b58\uff08flash memory\uff09\u6765\u5b58\u50a8\u548c\u6309\u9700\u52a0\u8f7d\u6a21\u578b\u53c2\u6570\u3002\u5176\u4e3b\u8981\u76ee\u6807\u662f\u89e3\u51b3\u5f53\u524d\u8bbe\u5907\u7684\u5185\u5b58\u5bb9\u91cf\u65e0\u6cd5\u4e00\u6b21\u6027\u52a0\u8f7d\u6574\u4e2a\u6a21\u578b\u7684\u95ee\u9898\uff0c\u4ece\u800c\u4f7f\u8fd9\u4e9b\u8bbe\u5907\u80fd\u591f\u5904\u7406\u8d85\u51fa\u5176\u5185\u5b58\u9650\u5236\u7684\u5927\u89c4\u6a21\u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aKeivan Alizadeh, Iman Mirzadeh, Dmitry Belenko, Karen Khatamifard, Minsik Cho, Carlo C Del Mundo, Mohammad Rastegari, Mehrdad Farajtabar\uff0c\u5747\u6765\u81eaApple\u516c\u53f8\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"509\" height=\"529\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-18.png\" alt=\"\" class=\"wp-image-4915\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-18.png 509w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-18-289x300.png 289w\" sizes=\"auto, (max-width: 509px) 100vw, 509px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2312.11514\" target=\"_blank\" rel=\"noreferrer noopener\">LLM in a flash: Efficient Large Language Model Inference with Limited Memory<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u4ee5\u4e0b\u662f\u5bf9\u8be5\u8bba\u6587\u8be6\u7ec6\u7684\u5206\u6790\u4e0e\u8bf4\u660e\uff1a<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1. <strong>\u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a<\/strong><\/h4>\n\n\n\n<p>\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08\u5982GPT\u7cfb\u5217\u3001LLaMA\u7b49\uff09\u5df2\u7ecf\u5728\u5404\u79cd\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u4e2d\u5c55\u793a\u4e86\u4f18\u5f02\u7684\u8868\u73b0\u3002\u7136\u800c\uff0c\u968f\u7740\u8fd9\u4e9b\u6a21\u578b\u7684\u89c4\u6a21\u4e0d\u65ad\u589e\u957f\uff0c\u5b83\u4eec\u5728\u63a8\u7406\u9636\u6bb5\u7684\u5185\u5b58\u9700\u6c42\u548c\u8ba1\u7b97\u91cf\u4e5f\u6210\u500d\u589e\u52a0\uff0c\u7279\u522b\u662f\u5f53\u6a21\u578b\u5305\u542b\u6570\u767e\u4ebf\u751a\u81f3\u4e0a\u4e07\u4ebf\u53c2\u6570\u65f6\uff0c\u5176\u9700\u8981\u7684\u5185\u5b58\u548c\u8ba1\u7b97\u80fd\u529b\u8fdc\u8d85\u51fa\u666e\u901a\u4e2a\u4eba\u8bbe\u5907\uff08\u5982\u667a\u80fd\u624b\u673a\u3001\u7b14\u8bb0\u672c\u7535\u8111\u7b49\uff09\u7684\u8d1f\u8f7d\u80fd\u529b\u3002<\/p>\n\n\n\n<p><strong>\u73b0\u72b6<\/strong>\uff1a\u4f20\u7edf\u7684\u63a8\u7406\u65b9\u6cd5\u4f9d\u8d56\u4e8e\u5c06\u6574\u4e2a\u6a21\u578b\u52a0\u8f7d\u5230DRAM\uff08\u52a8\u6001\u968f\u673a\u5b58\u53d6\u5b58\u50a8\u5668\uff09\u4e2d\u8fdb\u884c\u8ba1\u7b97\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u6709\u4ee5\u4e0b\u5c40\u9650\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bbe\u5907\u7684DRAM\u5bb9\u91cf\u5f80\u5f80\u6709\u9650\u3002\u4f8b\u5982\uff0c\u4e00\u4e2a7B\u53c2\u6570\u7684\u6a21\u578b\u4ec5\u52a0\u8f7d\u5176\u53c2\u6570\u5c31\u9700\u8981\u8d85\u8fc714GB\u7684\u5185\u5b58\uff0c\u800c\u5927\u591a\u6570\u667a\u80fd\u624b\u673a\u7684\u5185\u5b58\u8fdc\u4f4e\u4e8e\u6b64\u3002<\/li>\n\n\n\n<li>\u4f7f\u7528\u8bf8\u5982\u91cf\u5316\uff08quantization\uff09\u4e4b\u7c7b\u7684\u6280\u672f\u867d\u7136\u53ef\u4ee5\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u51cf\u5c11\u6a21\u578b\u5927\u5c0f\uff0c\u4f46\u4ecd\u65e0\u6cd5\u7a81\u7834\u9700\u8981\u5c06\u6574\u4e2a\u6a21\u578b\u52a0\u8f7d\u5230DRAM\u4e2d\u7684\u786c\u6027\u8981\u6c42\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>\u521b\u65b0\u52a8\u673a<\/strong>\uff1a\u4e3a\u4e86\u514b\u670d\u4e0a\u8ff0\u5c40\u9650\uff0c\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u5373\u5c06\u6a21\u578b\u53c2\u6570\u5b58\u50a8\u5728<strong>\u95ea\u5b58<\/strong>\u4e2d\uff0c\u5e76\u6839\u636e\u63a8\u7406\u7684\u9700\u8981\u6309\u9700\u5c06\u53c2\u6570\u52a0\u8f7d\u5230DRAM\u4e2d\u3002\u8fd9\u79cd\u65b9\u6cd5\u80fd\u591f\u907f\u514d\u4e00\u6b21\u6027\u5c06\u6574\u4e2a\u6a21\u578b\u52a0\u8f7d\u5230\u5185\u5b58\u4e2d\uff0c\u4ece\u800c\u5927\u5e45\u51cf\u5c11\u5185\u5b58\u7684\u9700\u6c42\uff0c\u751a\u81f3\u53ef\u4ee5\u5728\u8bbe\u5907\u5185\u5b58\u53ea\u6709\u6a21\u578b\u5927\u5c0f\u4e00\u534a\u751a\u81f3\u66f4\u5c11\u7684\u60c5\u51b5\u4e0b\u8fd0\u884c\u5927\u6a21\u578b\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">2. <strong>\u95ee\u9898\u4e0e\u6311\u6218<\/strong><\/h4>\n\n\n\n<p>\u8bba\u6587\u4e2d\u6307\u51fa\uff0c\u867d\u7136\u95ea\u5b58\u7684\u5b58\u50a8\u5bb9\u91cf\u6bd4DRAM\u5927\u5f97\u591a\uff0c\u4f46\u5176\u8bfb\u5199\u6027\u80fd\uff08\u5c24\u5176\u662f\u968f\u673a\u8bbf\u95ee\u7684\u6027\u80fd\uff09\u8fdc\u4f4e\u4e8eDRAM\u3002\u95ea\u5b58\u548cDRAM\u4e4b\u95f4\u7684\u6027\u80fd\u5dee\u5f02\u4e3b\u8981\u4f53\u73b0\u5728\u4ee5\u4e0b\u65b9\u9762\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5e26\u5bbd<\/strong>\uff1aDRAM\u7684\u5e26\u5bbd\u901a\u5e38\u662f100GB\/s\u7ea7\u522b\uff0c\u800c\u95ea\u5b58\u7684\u5e26\u5bbd\u5927\u7ea6\u53ea\u67091GB\/s\u3002\u8fd9\u610f\u5473\u7740\u4ece\u95ea\u5b58\u4e2d\u8bfb\u53d6\u6570\u636e\u7684\u901f\u5ea6\u8fdc\u4f4e\u4e8eDRAM\u3002<\/li>\n\n\n\n<li><strong>\u968f\u673a\u8bfb\u53d6\u5ef6\u8fdf<\/strong>\uff1a\u95ea\u5b58\u7684\u968f\u673a\u8bfb\u53d6\uff08\u5c24\u5176\u662f\u5c0f\u5757\u6570\u636e\u7684\u8bfb\u53d6\uff09\u7531\u4e8e\u591a\u9636\u6bb5\u7684\u8bbf\u95ee\u8fc7\u7a0b\uff08\u6d89\u53ca\u64cd\u4f5c\u7cfb\u7edf\u3001\u9a71\u52a8\u7a0b\u5e8f\u3001\u4e2d\u65ad\u5904\u7406\u3001\u95ea\u5b58\u63a7\u5236\u5668\u7b49\uff09\uff0c\u5bfc\u81f4\u5176\u8bfb\u53d6\u5ef6\u8fdf\u8f83\u957f\u3002\u8fd9\u4e9b\u9636\u6bb5\u7684\u5ef6\u8fdf\u5728\u5904\u7406\u5c0f\u5757\u6570\u636e\u65f6\u663e\u5f97\u5c24\u4e3a\u7a81\u51fa\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u56e0\u6b64\uff0c\u7b80\u5355\u5730\u5c06\u6a21\u578b\u5b58\u50a8\u5728\u95ea\u5b58\u4e2d\u5e76\u6309\u9700\u52a0\u8f7d\u5e76\u4e0d\u8db3\u4ee5\u89e3\u51b3\u95ee\u9898\uff0c\u8bba\u6587\u4f5c\u8005\u63d0\u51fa\u4e86\u4e00\u7cfb\u5217\u7684\u4f18\u5316\u7b56\u7565\uff0c\u65e8\u5728\u51cf\u5c11\u4ece\u95ea\u5b58\u8bfb\u53d6\u6570\u636e\u7684\u91cf\u5e76\u63d0\u9ad8\u8bfb\u53d6\u7684\u6548\u7387\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">3. <strong>\u6838\u5fc3\u65b9\u6cd5\u4e0e\u6280\u672f\u521b\u65b0<\/strong><\/h4>\n\n\n\n<p>\u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u8bba\u6587\u63d0\u51fa\u4e86\u4e24\u4e2a\u6838\u5fc3\u6280\u672f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7a97\u53e3\u6280\u672f\uff08Windowing\uff09<\/strong>\uff1a\u901a\u8fc7\u91cd\u7528\u5148\u524d\u5df2\u6fc0\u6d3b\u7684\u795e\u7ecf\u5143\uff0c\u51cf\u5c11\u4ece\u95ea\u5b58\u4f20\u8f93\u7684\u6570\u636e\u91cf\uff0c\u4ece\u800c\u964d\u4f4eI\/O\u5f00\u9500\u3002<\/li>\n\n\n\n<li><strong>\u884c\u5217\u7ed1\u5b9a\uff08Row-Column Bundling\uff09<\/strong>\uff1a\u4e3a\u4e86\u5145\u5206\u5229\u7528\u95ea\u5b58\u7684\u987a\u5e8f\u8bbf\u95ee\u4f18\u52bf\uff0c\u4f5c\u8005\u63d0\u51fa\u5c06\u77e9\u9635\u4e2d\u7684\u884c\u548c\u5217\u7ed1\u5b9a\u5728\u4e00\u8d77\u8bfb\u53d6\uff0c\u4ee5\u6b64\u589e\u5927\u6570\u636e\u8bfb\u53d6\u5757\u7684\u5927\u5c0f\uff0c\u4ece\u800c\u63d0\u9ad8\u8bfb\u53d6\u6548\u7387\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">3.1 <strong>\u7a97\u53e3\u6280\u672f<\/strong><\/h5>\n\n\n\n<p>\u7a97\u53e3\u6280\u672f\u7684\u4e3b\u8981\u76ee\u7684\u662f\u901a\u8fc7\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u91cd\u7528\u5148\u524d\u7684\u795e\u7ecf\u5143\u6fc0\u6d3b\u503c\u6765\u51cf\u5c11\u4ece\u95ea\u5b58\u52a0\u8f7d\u7684\u6570\u636e\u91cf\u3002\u5177\u4f53\u65b9\u6cd5\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u795e\u7ecf\u5143\u6fc0\u6d3b\u7a00\u758f\u6027<\/strong>\uff1a\u7814\u7a76\u53d1\u73b0\uff0c\u8bb8\u591a\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u5c42\uff08FFN\uff09\u5177\u6709\u9ad8\u5ea6\u7a00\u758f\u6027\u3002\u6bd4\u5982\u5728OPT-6.7B\u6a21\u578b\u4e2d\uff0cFFN\u5c42\u7684\u6fc0\u6d3b\u7a00\u758f\u6027\u53ef\u4ee5\u8fbe\u523097%\uff0c\u5373\u5728\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\uff0c\u5927\u591a\u6570\u795e\u7ecf\u5143\u662f\u672a\u88ab\u6fc0\u6d3b\u7684\u3002\u8fd9\u610f\u5473\u7740\u5728\u63a8\u7406\u65f6\uff0c\u5e76\u4e0d\u9700\u8981\u52a0\u8f7d\u6240\u6709\u7684\u6a21\u578b\u53c2\u6570\uff0c\u800c\u53ea\u9700\u52a0\u8f7d\u90a3\u4e9b\u88ab\u6fc0\u6d3b\u7684\u795e\u7ecf\u5143\u76f8\u5173\u7684\u53c2\u6570\u3002<\/li>\n\n\n\n<li><strong>\u6ed1\u52a8\u7a97\u53e3\u673a\u5236<\/strong>\uff1a\u4f5c\u8005\u63d0\u51fa\u4e86\u4e00\u4e2a\u201c\u6ed1\u52a8\u7a97\u53e3\u201d\u7b56\u7565\uff0c\u5373\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u53ea\u4fdd\u7559\u6700\u8fd1\u8f93\u5165\u7684\u90e8\u5206token\u6240\u9700\u8981\u7684\u795e\u7ecf\u5143\u6743\u91cd\u6570\u636e\u3002\u5bf9\u4e8e\u65b0\u7684\u8f93\u5165token\uff0c\u53ea\u9700\u589e\u91cf\u5730\u52a0\u8f7d\u4e0e\u4e4b\u524d\u4e0d\u540c\u7684\u795e\u7ecf\u5143\u6570\u636e\u5373\u53ef\u3002\u8fd9\u6837\u53ef\u4ee5\u907f\u514d\u91cd\u590d\u52a0\u8f7d\u4e0d\u5fc5\u8981\u7684\u6743\u91cd\uff0c\u4ece\u800c\u63d0\u9ad8\u5185\u5b58\u5229\u7528\u7387\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">3.2 <strong>\u884c\u5217\u7ed1\u5b9a\u6280\u672f<\/strong><\/h5>\n\n\n\n<p>\u884c\u5217\u7ed1\u5b9a\u7684\u601d\u8def\u662f\u57fa\u4e8eFFN\u5c42\u4e2d\u884c\u548c\u5217\u7684\u6743\u91cd\u8bbf\u95ee\u662f\u5bc6\u5207\u76f8\u5173\u7684\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u884c\u5217\u5bf9\u5e94\u6027<\/strong>\uff1a\u5728\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u5c42\u4e2d\uff0c\u6743\u91cd\u77e9\u9635\u7684\u5217\u5bf9\u5e94\u4e8e\u4e0a\u6295\u5f71\uff08up projection\uff09\u64cd\u4f5c\uff0c\u800c\u884c\u5bf9\u5e94\u4e8e\u4e0b\u6295\u5f71\uff08down projection\uff09\u64cd\u4f5c\u3002\u7531\u4e8e\u8fd9\u4e24\u4e2a\u64cd\u4f5c\u4e2d\u7684\u7b2ci\u5217\u548c\u7b2ci\u884c\u662f\u540c\u65f6\u88ab\u4f7f\u7528\u7684\uff0c\u56e0\u6b64\u53ef\u4ee5\u5c06\u5b83\u4eec\u7ed1\u5b9a\u5728\u4e00\u8d77\u5b58\u50a8\u5e76\u8bfb\u53d6\u3002<\/li>\n\n\n\n<li><strong>\u7ed1\u5b9a\u4f18\u52bf<\/strong>\uff1a\u901a\u8fc7\u7ed1\u5b9a\u8fd9\u4e9b\u5bf9\u5e94\u7684\u884c\u548c\u5217\uff0c\u4f5c\u8005\u80fd\u591f\u5c06\u6570\u636e\u6253\u5305\u6210\u66f4\u5927\u7684\u5757\u8fdb\u884c\u8bfb\u53d6\uff0c\u4ece\u800c\u51cf\u5c11\u591a\u6b21\u5c0f\u5757\u8bfb\u53d6\u5e26\u6765\u7684\u5ef6\u8fdf\uff0c\u63d0\u9ad8\u4e86\u603b\u7684\u8bfb\u53d6\u541e\u5410\u91cf\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">4. <strong>\u5b9e\u9a8c\u4e0e\u7ed3\u679c\u5206\u6790<\/strong><\/h4>\n\n\n\n<p>\u8bba\u6587\u901a\u8fc7\u4e00\u7cfb\u5217\u5b9e\u9a8c\u9a8c\u8bc1\u4e86\u5176\u65b9\u6cd5\u7684\u6709\u6548\u6027\u3002\u5b9e\u9a8c\u4e3b\u8981\u9488\u5bf9\u51e0\u79cd\u4e0d\u540c\u89c4\u6a21\u7684\u6a21\u578b\uff08\u5305\u62ecOPT-6.7B\u3001Falcon 7B\u3001Persimmon 8B\u548cLlama 2\uff09\u8fdb\u884c\u8bc4\u4f30\uff0c\u6d4b\u8bd5\u5e73\u53f0\u5305\u62ec\u82f9\u679cM1 Max\u3001M2 Ultra\u4ee5\u53ca\u914d\u5907NVIDIA RTX 4090\u7684Linux\u8bbe\u5907\u3002\u5173\u952e\u5b9e\u9a8c\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u63a8\u7406\u901f\u5ea6\u7684\u63d0\u5347<\/strong>\uff1a\u8bba\u6587\u663e\u793a\uff0c\u5728DRAM\u5bb9\u91cf\u53ea\u6709\u6a21\u578b\u4e00\u534a\u7684\u60c5\u51b5\u4e0b\uff0c\u4f7f\u7528\u6240\u63d0\u51fa\u7684\u6280\u672f\uff0c\u63a8\u7406\u901f\u5ea6\u5728CPU\u4e0a\u63d0\u5347\u4e864\u500d\uff0c\u5728GPU\u4e0a\u63d0\u5347\u4e8620\u500d\u3002\u5728GPU\u4e0a\u4f7f\u7528\u91d1\u5c5e\uff08Metal\uff09\u6846\u67b6\u8fd0\u884c\u65f6\uff0c\u63a8\u7406\u901f\u5ea6\u751a\u81f3\u63d0\u5347\u4e8625\u500d\u3002<\/li>\n\n\n\n<li><strong>\u6570\u636e\u4f20\u8f93\u91cf\u7684\u51cf\u5c11<\/strong>\uff1a\u901a\u8fc7\u9884\u6d4b\u5668\uff08predictor\uff09\u548c\u7a97\u53e3\u6280\u672f\uff0c\u4f5c\u8005\u80fd\u591f\u5c06\u4ece\u95ea\u5b58\u52a0\u8f7d\u7684\u6570\u636e\u91cf\u51cf\u5c11\u81f3\u539f\u6765\u76841\/65\u3002\u5728DRAM\u6709\u9650\u7684\u60c5\u51b5\u4e0b\uff0c\u7cfb\u7edf\u80fd\u591f\u4ec5\u52a0\u8f7d\u5fc5\u8981\u7684\u6743\u91cd\uff0c\u4ece\u800c\u51cf\u5c11I\/O\u5ef6\u8fdf\u3002<\/li>\n\n\n\n<li><strong>\u7aef\u5230\u7aef\u63a8\u7406\u5ef6\u8fdf\u7684\u51cf\u5c11<\/strong>\uff1a\u5b9e\u9a8c\u8868\u660e\uff0c\u5728CPU\u548cGPU\u4e0a\uff0c\u8be5\u65b9\u6cd5\u5728\u7aef\u5230\u7aef\u7684\u63a8\u7406\u8fc7\u7a0b\u4e2d\u663e\u8457\u51cf\u5c11\u4e86I\/O\u548c\u8ba1\u7b97\u5ef6\u8fdf\u3002\u4f8b\u5982\uff0cOPT-6.7B\u6a21\u578b\u7684\u603b\u63a8\u7406\u5ef6\u8fdf\u4ece\u539f\u6765\u76843182ms\u51cf\u5c11\u5230669ms\uff0c\u663e\u8457\u63d0\u5347\u4e86\u63a8\u7406\u6548\u7387\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">5. <strong>\u4e0e\u5176\u4ed6\u5de5\u4f5c\u6bd4\u8f83<\/strong><\/h4>\n\n\n\n<p>\u8bba\u6587\u5c06\u5176\u65b9\u6cd5\u4e0e\u73b0\u6709\u7684\u6a21\u578b\u538b\u7f29\u6280\u672f\u548c\u7a00\u758f\u6027\u4f18\u5316\u65b9\u6cd5\u8fdb\u884c\u4e86\u6bd4\u8f83\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6a21\u578b\u538b\u7f29<\/strong>\uff1a\u867d\u7136\u6a21\u578b\u538b\u7f29\uff08\u5982\u91cf\u5316\u3001\u526a\u679d\u7b49\uff09\u53ef\u4ee5\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u51cf\u5c0f\u6a21\u578b\u7684\u5927\u5c0f\uff0c\u4f46\u8fd9\u4e9b\u65b9\u6cd5\u4ecd\u9700\u5c06\u6574\u4e2a\u6a21\u578b\u52a0\u8f7d\u5230DRAM\u4e2d\uff0c\u65e0\u6cd5\u89e3\u51b3\u5185\u5b58\u53d7\u9650\u8bbe\u5907\u7684\u95ee\u9898\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u8bba\u6587\u63d0\u51fa\u7684\u6309\u9700\u52a0\u8f7d\u65b9\u6cd5\u53ef\u4ee5\u5728\u4e0d\u538b\u7f29\u6a21\u578b\u7684\u60c5\u51b5\u4e0b\u5904\u7406\u5927\u6a21\u578b\uff0c\u7279\u522b\u9002\u7528\u4e8e\u5185\u5b58\u975e\u5e38\u6709\u9650\u7684\u8bbe\u5907\u3002<\/li>\n\n\n\n<li><strong>\u9009\u62e9\u6027\u6743\u91cd\u52a0\u8f7d<\/strong>\uff1a\u4f5c\u8005\u63d0\u5230\u73b0\u6709\u7684Dejavu\u65b9\u6cd5\u867d\u7136\u4e5f\u5229\u7528\u4e86\u6fc0\u6d3b\u7a00\u758f\u6027\uff0c\u4f46\u5b83\u4f9d\u7136\u8981\u6c42\u4eceGPU\u5185\u5b58\u52a0\u8f7d\u6743\u91cd\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u672c\u6587\u65b9\u6cd5\u53ef\u4ee5\u5728GPU\u548cDRAM\u90fd\u65e0\u6cd5\u5bb9\u7eb3\u6574\u4e2a\u6a21\u578b\u7684\u60c5\u51b5\u4e0b\u4f7f\u7528\u95ea\u5b58\u5b58\u50a8\uff0c\u5e76\u4e14\u901a\u8fc7\u4f18\u5316\u6570\u636e\u52a0\u8f7d\u7684\u7b56\u7565\u8fdb\u4e00\u6b65\u51cf\u5c11I\/O\u5ef6\u8fdf\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">6. <strong>\u65b9\u6cd5\u7684\u5b9e\u9645\u5e94\u7528\u4e0e\u610f\u4e49<\/strong><\/h4>\n\n\n\n<p>\u8be5\u7814\u7a76\u7684\u5b9e\u9645\u610f\u4e49\u5728\u4e8e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8fb9\u7f18\u8bbe\u5907\u7684\u6a21\u578b\u63a8\u7406<\/strong>\uff1a\u901a\u8fc7\u8be5\u65b9\u6cd5\uff0c\u8fb9\u7f18\u8bbe\u5907\uff08\u5982\u667a\u80fd\u624b\u673a\u3001\u5d4c\u5165\u5f0f\u8bbe\u5907\u7b49\uff09\u53ef\u4ee5\u8fd0\u884c\u6bd4\u5176\u5185\u5b58\u5bb9\u91cf\u5927\u4e24\u500d\u7684\u6a21\u578b\uff0c\u4ece\u800c\u5927\u5927\u6269\u5c55\u4e86\u8fd9\u4e9b\u8bbe\u5907\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u667a\u80fd\u52a9\u624b\u7b49\u5e94\u7528\u4e2d\u7684\u80fd\u529b\u3002<\/li>\n\n\n\n<li><strong>\u786c\u4ef6\u4f18\u5316\u7684\u63a8\u7406\u6210\u672c\u6a21\u578b<\/strong>\uff1a\u4f5c\u8005\u901a\u8fc7\u6df1\u5165\u7406\u89e3\u786c\u4ef6\u7279\u6027\uff08\u95ea\u5b58\u548cDRAM\u7684\u5e26\u5bbd\u548c\u5ef6\u8fdf\uff09\u6784\u5efa\u4e86\u4e00\u4e2a\u63a8\u7406\u6210\u672c\u6a21\u578b\uff0c\u5e76\u63d0\u51fa\u4e86\u4e24\u79cd\u4f18\u5316\u6280\u672f\uff0c\u4f7f\u5f97\u63a8\u7406\u8fc7\u7a0b\u4e2d\u80fd\u591f\u6709\u6548\u51cf\u5c11\u4ece\u95ea\u5b58\u52a0\u8f7d\u7684\u6570\u636e\u91cf\uff0c\u63d0\u9ad8\u8bfb\u53d6\u541e\u5410\u91cf\uff0c\u4ece\u800c\u52a0\u901f\u6a21\u578b\u63a8\u7406\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">7. <strong>\u5c40\u9650\u4e0e\u672a\u6765\u7814\u7a76\u65b9\u5411<\/strong><\/h4>\n\n\n\n<p>\u5c3d\u7ba1\u8be5\u8bba\u6587\u5c55\u793a\u4e86\u5728\u5185\u5b58\u53d7\u9650\u8bbe\u5907\u4e0a\u8fd0\u884c\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u7684\u663e\u8457\u4f18\u52bf\uff0c\u4f46\u4f5c\u8005\u4e5f\u6307\u51fa\u4e86\u4e00\u4e9b\u5c40\u9650\u548c\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u529f\u8017\u4e0e\u70ed\u91cf\u9650\u5236<\/strong>\uff1a\u7531\u4e8e\u8bbe\u5907\u5728\u957f\u65f6\u95f4\u7684\u63a8\u7406\u8fc7\u7a0b\u4e2d\u4f1a\u9762\u4e34\u529f\u8017\u548c\u70ed\u91cf\u95ee\u9898\uff0c\u672a\u6765\u7684\u7814\u7a76\u53ef\u4ee5\u8fdb\u4e00\u6b65\u4f18\u5316\u8fd9\u4e9b\u65b9\u9762\u3002<\/li>\n\n\n\n<li><strong>\u6279\u5904\u7406\u63a8\u7406<\/strong>\uff1a\u76ee\u524d\u7684\u7814\u7a76\u4e3b\u8981\u9488\u5bf9\u5355\u5e8f\u5217\u7684\u63a8\u7406\uff0c\u800c\u591a\u6279\u6b21\u63a8\u7406\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u66f4\u4e3a\u5e38\u89c1\uff0c\u5982\u4f55\u5728\u6279\u5904\u7406\u63a8\u7406\u4e2d\u8fdb\u4e00\u6b65\u4f18\u5316\u5c06\u662f\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\u4e4b\u4e00\u3002<\/li>\n\n\n\n<li><strong>\u975e\u7a00\u758f\u7f51\u7edc\u7684\u6743\u91cd\u9009\u62e9\u6027\u52a0\u8f7d<\/strong>\uff1a\u867d\u7136\u672c\u6587\u7684\u6280\u672f\u4e3b\u8981\u9488\u5bf9\u7a00\u758f\u7f51\u7edc\uff0c\u4f46\u5176\u6838\u5fc3\u601d\u60f3\u4e5f\u53ef\u4ee5\u6269\u5c55\u5230\u975e\u7a00\u758f\u7f51\u7edc\uff0c\u5373\u6839\u636e\u8f93\u5165\u6216\u4e0a\u4e0b\u6587\u9009\u62e9\u6027\u5730\u52a0\u8f7d\u6743\u91cd\u3002\u672a\u6765\u7684\u5de5\u4f5c\u53ef\u4ee5\u63a2\u7d22\u8fd9\u4e00\u601d\u8def\u5728\u66f4\u5e7f\u6cdb\u5e94\u7528\u4e2d\u7684\u6f5c\u529b\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">8. <strong>\u603b\u7ed3<\/strong><\/h4>\n\n\n\n<p>\u8be5\u8bba\u6587\u4e3a\u5982\u4f55\u5728\u5185\u5b58\u53d7\u9650\u8bbe\u5907\u4e0a\u9ad8\u6548\u8fd0\u884c\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u63d0\u4f9b\u4e86\u5168\u65b0\u7684\u89e3\u51b3\u65b9\u6848\u3002\u901a\u8fc7\u95ea\u5b58\u548cDRAM\u4e4b\u95f4\u7684\u6570\u636e\u4f18\u5316\u52a0\u8f7d\u7b56\u7565\uff0c\u4f5c\u8005\u63d0\u51fa\u7684\u65b9\u6cd5\u5728\u4e0d\u538b\u7f29\u6a21\u578b\u7684\u60c5\u51b5\u4e0b\u5b9e\u73b0\u4e86\u663e\u8457\u7684\u63a8\u7406\u52a0\u901f\uff0c\u4e3a\u5927\u6a21\u578b\u7684\u666e\u53ca\u548c\u5e94\u7528\u63d0\u4f9b\u4e86\u66f4\u5e7f\u9614\u7684\u524d\u666f\u3002\u8fd9\u9879\u7814\u7a76\u5c55\u793a\u4e86\u786c\u4ef6\u5bfc\u5411\u7684\u673a\u5668\u5b66\u4e60\u4f18\u5316\u7b56\u7565\u5728\u8fb9\u7f18\u8ba1\u7b97\u8bbe\u5907\u4e0a\u7684\u5de8\u5927\u6f5c\u529b\uff0c\u5e76\u4e3a\u672a\u6765\u5728\u6709\u9650\u8d44\u6e90\u8bbe\u5907\u4e0a\u8fd0\u884c\u66f4\u5927\u89c4\u6a21\u6a21\u578b\u63d0\u4f9b\u4e86\u6709\u529b\u7684\u6280\u672f\u652f\u6301\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587LLM in a flash: Efficient Large Language Model Infere [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","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":"","ast-disable-related-posts":"","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":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"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":""},"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":[8],"tags":[39,86,85],"class_list":["post-4914","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-llm","tag-apple"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587LLM in a flash: Efficient Large Language Model Infere&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4914","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4914"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4914\/revisions"}],"predecessor-version":[{"id":4916,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4914\/revisions\/4916"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4914"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4914"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4914"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}