{"id":2535,"date":"2024-03-14T15:14:24","date_gmt":"2024-03-14T07:14:24","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=2535"},"modified":"2024-03-14T15:21:49","modified_gmt":"2024-03-14T07:21:49","slug":"%e7%94%a8-bitsandbytes%e3%80%814-%e6%af%94%e7%89%b9%e9%87%8f%e5%8c%96%e5%92%8c-qlora-%e6%89%93%e9%80%a0%e4%ba%b2%e6%b0%91%e7%9a%84-llm","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=2535","title":{"rendered":"\u7528 bitsandbytes\u30014 \u6bd4\u7279\u91cf\u5316\u548c QLoRA \u6253\u9020\u4eb2\u6c11\u7684 LLM"},"content":{"rendered":"\n<figure class=\"wp-block-image\"><a href=\"https:\/\/www.zhihu.com\/org\/huggingface\" target=\"_blank\" rel=\"noreferrer noopener\"><img loading=\"lazy\" decoding=\"async\" width=\"160\" height=\"160\" src=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image.png\" alt=\"Hugging Face\" class=\"wp-image-2536\" srcset=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image.png 160w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-150x150.png 150w\" sizes=\"auto, (max-width: 160px) 100vw, 160px\" \/><\/a><\/figure>\n\n\n\n<p><a href=\"https:\/\/www.zhihu.com\/org\/huggingface\" target=\"_blank\" rel=\"noreferrer noopener\">Hugging Face<\/a><a href=\"https:\/\/www.zhihu.com\/question\/48510028\" target=\"_blank\" rel=\"noreferrer noopener\">\u200b<\/a><\/p>\n\n\n\n<p>\u5df2\u8ba4\u8bc1\u8d26\u53f7<\/p>\n\n\n\n<p>47 \u4eba\u8d5e\u540c\u4e86\u8be5\u6587\u7ae0<\/p>\n\n\n\n<p>\u4f17\u6240\u5468\u77e5\uff0cLLM \u89c4\u6a21\u5e9e\u5927\uff0c\u5982\u679c\u5728\u4e5f\u80fd\u6d88\u8d39\u7c7b\u786c\u4ef6\u4e2d\u8fd0\u884c\u6216\u8bad\u7ec3\u5b83\u4eec\u5c06\u662f\u5176\u4eb2\u6c11\u5316\u7684\u5de8\u5927\u8fdb\u6b65\u3002\u6211\u4eec\u4e4b\u524d\u64b0\u5199\u7684&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/huggingface.co\/blog\/zh\/hf-bitsandbytes-integration\" target=\"_blank\" rel=\"noreferrer noopener\">LLM.int8 \u535a\u6587<\/a><\/strong>&nbsp;\u5c55\u793a\u4e86\u6211\u4eec\u662f\u5982\u4f55\u5c06&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/arxiv.org\/abs\/2208.07339\" target=\"_blank\" rel=\"noreferrer noopener\">LLM.int8 \u8bba\u6587<\/a><\/strong>&nbsp;\u4e2d\u7684\u6280\u672f\u901a\u8fc7&nbsp;<code>bitsandbytes<\/code>&nbsp;\u5e93\u96c6\u6210\u5230&nbsp;<code>transformers<\/code>&nbsp;\u4e2d\u7684\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u6211\u4eec\u4e0d\u65ad\u52aa\u529b\u4ee5\u4e0d\u65ad\u964d\u4f4e\u5927\u6a21\u578b\u7684\u51c6\u5165\u95e8\u69db\u3002\u5728\u6b64\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u51b3\u5b9a\u518d\u6b21\u4e0e&nbsp;<code>bitsandbytes<\/code>&nbsp;\u8054\u624b\uff0c\u652f\u6301\u7528\u6237\u4ee5 4 \u6bd4\u7279\u7cbe\u5ea6\u8fd0\u884c\u4efb\u4f55\u6a21\u6001 (\u6587\u672c\u3001\u89c6\u89c9\u3001\u591a\u6a21\u6001\u7b49) \u4e0a\u7684\u7edd\u5927\u591a\u6570 HF \u6a21\u578b\u3002\u7528\u6237\u8fd8\u53ef\u4ee5\u5229\u7528 Hugging Face \u751f\u6001\u7cfb\u7edf\u4e2d\u7684\u5de5\u5177\u5728 4 \u6bd4\u7279\u6a21\u578b\u4e4b\u4e0a\u8bad\u7ec3\u9002\u914d\u5668\u3002\u8fd9\u4e00\u5de5\u4f5c\u57fa\u4e8e Dettmers \u7b49\u4eba\u6700\u8fd1\u5728 QLoRA \u8fd9\u7bc7\u8bba\u6587\u4e2d\u4ecb\u7ecd\u7684\u4e00\u79cd\u65b0\u65b9\u6cd5\uff0c\u5176\u8bba\u6587\u6458\u8981\u5982\u4e0b:<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u6211\u4eec\u63d0\u51fa\u4e86 QLoRA\uff0c\u8fd9\u662f\u4e00\u79cd\u9ad8\u6548\u7684\u5fae\u8c03\u65b9\u6cd5\uff0c\u53ef\u51cf\u5c11\u5185\u5b58\u4f7f\u7528\u91cf\uff0c\u4f7f\u5f97\u5728\u5355\u4e2a 48GB GPU \u4e0a\u5c31\u53ef\u4ee5\u5fae\u8c03 65B \u7684\u6a21\u578b\uff0c\u800c\u4e14\u6240\u5f97\u6a21\u578b\u7684\u6027\u80fd\u4e0e\u5168 16 \u6bd4\u7279\u5fae\u8c03\u76f8\u5f53\u3002QLoRA \u901a\u8fc7\u51bb\u7ed3 4 \u6bd4\u7279\u91cf\u5316\u7684\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5c06\u68af\u5ea6\u53cd\u5411\u4f20\u64ad\u5230\u4f4e\u79e9\u9002\u914d\u5668 (LoRA) \u4e2d\u3002\u6211\u4eec\u6700\u597d\u7684\u6a21\u578b (\u6211\u4eec\u5c06\u5176\u547d\u540d\u4e3a Guanaco) \u4ec5\u9700\u5728\u5355\u4e2a GPU \u4e0a\u8fdb\u884c 24 \u5c0f\u65f6\u5fae\u8c03\uff0c\u5c31\u80fd\u5728 Vicuna \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u4f18\u4e8e\u6240\u6709\u4e4b\u524d\u516c\u5f00\u53d1\u5e03\u7684\u6a21\u578b\uff0c\u4e14\u8fbe\u5230\u4e86 ChatGPT \u6027\u80fd\u6c34\u5e73\u7684 99.3%\u3002QLoRA \u5f15\u5165\u4e86\u591a\u9879\u521b\u65b0\u6280\u672f\uff0c\u5728\u4e0d\u727a\u7272\u6027\u80fd\u7684\u60c5\u51b5\u4e0b\u8282\u7701\u5185\u5b58:(a) 4 \u4f4d NormalFloat (NF4)\uff0c\u4e00\u79cd\u65b0\u7684\u6570\u636e\u7c7b\u578b\uff0c\u5728\u4fe1\u606f\u8bba\u610f\u4e49\u4e0a\u662f\u6b63\u6001\u5206\u5e03\u6743\u91cd\u7684\u6700\u4f73\u8868\u793a (b) \u53cc\u91cf\u5316\uff0c\u901a\u8fc7\u5bf9\u91cf\u5316\u7cfb\u6570\u8fdb\u884c\u4e8c\u6b21\u91cf\u5316\u6765\u51cf\u5c11\u5e73\u5747\u5185\u5b58\u5360\u7528\uff0c\u4ee5\u53ca (c) \u7528\u4e8e\u964d\u4f4e\u5cf0\u503c\u5185\u5b58\u5360\u7528\u7684\u5206\u9875\u4f18\u5316\u5668\u3002\u6211\u4eec\u4f7f\u7528 QLoRA \u5fae\u8c03\u4e86 1000 \u591a\u4e2a\u6a21\u578b\uff0c\u5e76\u7ed9\u51fa\u4e86\u5b83\u4eec\u5728\u6307\u4ee4\u4f9d\u4ece\u3001\u804a\u5929\u7b49\u4efb\u52a1\u4e0a\u7684\u8be6\u7ec6\u6027\u80fd\u5206\u6790\uff0c\u5176\u4e2d\u6db5\u76d6\u4e86 8 \u4e2a\u6307\u4ee4\u6570\u636e\u96c6\u3001\u591a\u79cd\u6a21\u578b\u67b6\u6784 (LLaMA\u3001T5)\uff0c\u8fd8\u5305\u62ec\u4e86\u65e0\u6cd5\u7528\u5e38\u89c4\u65b9\u6cd5\u5fae\u8c03\u7684\u5927\u6a21\u578b (\u4f8b\u5982 33B \u548c 65B \u6a21\u578b)\u3002\u7ed3\u679c\u8868\u660e\uff0c\u5728\u5c0f\u578b\u9ad8\u8d28\u91cf\u6570\u636e\u96c6\u7684\u8fdb\u884c QLoRA \u5fae\u8c03\u80fd\u5e26\u6765\u6700\u5148\u8fdb\u7684\u6027\u80fd\uff0c\u4e14\u6240\u9700\u7684\u6a21\u578b\u5c3a\u5bf8\u66f4\u5c0f\u3002\u6211\u4eec\u4f7f\u7528\u4eba\u7c7b\u548c GPT-4 \u5bf9\u804a\u5929\u673a\u5668\u4eba\u7684\u6027\u80fd\u8fdb\u884c\u4e86\u8be6\u7ec6\u8bc4\u4f30\u5206\u6790\uff0c\u7ed3\u679c\u8868\u660e GPT-4 \u8bc4\u4f30\u662f\u66ff\u4ee3\u4eba\u7c7b\u8bc4\u4f30\u7684\u5ec9\u4ef7\u4e14\u5408\u7406\u7684\u65b9\u6848\u3002\u6b64\u5916\uff0c\u6211\u4eec\u53d1\u73b0\u5f53\u524d\u7684\u804a\u5929\u673a\u5668\u4eba\u57fa\u51c6\u6d4b\u8bd5\u5728\u51c6\u786e\u8bc4\u4f30\u804a\u5929\u673a\u5668\u4eba\u7684\u6027\u80fd\u6c34\u5e73\u8fd9\u4e00\u65b9\u9762\u5e76\u4e0d\u5341\u5206\u53ef\u4fe1\u3002\u6211\u4eec\u8fd8\u6311\u9009\u4e86\u4e00\u4e9b\u6837\u672c\uff0c\u5bf9 Guanaco \u6bd4 ChatGPT \u505a\u5f97\u4e0d\u597d\u7684\u6848\u4f8b\u8fdb\u884c\u4e86\u5206\u6790\u3002\u6211\u4eec\u53d1\u5e03\u4e86\u6240\u6709\u6a21\u578b\u548c\u4ee3\u7801\uff0c\u5305\u62ec\u7528\u4e8e 4 \u6bd4\u7279\u8bad\u7ec3\u7684 CUDA \u6838\u51fd\u6570\u3002<\/p>\n<\/blockquote>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u8d44\u6e90<\/strong><\/h2>\n\n\n\n<p>\u4e0b\u9762\u662f\u4e00\u4e9b 4 \u6bd4\u7279\u6a21\u578b\u548c QLoRA \u7684\u5165\u95e8\u8d44\u6e90:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/arxiv.org\/abs\/2305.14314\" target=\"_blank\" rel=\"noreferrer noopener\">\u539f\u59cb\u8bba\u6587<\/a><\/strong><\/li>\n\n\n\n<li><strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/colab.research.google.com\/drive\/1ge2F1QSK8Q7h0hn3YKuBCOAS0bK8E0wf%3Fusp%3Dsharing\" target=\"_blank\" rel=\"noreferrer noopener\">\u6709\u5173 bitsandbytes \u57fa\u7840\u7528\u6cd5\u7684 Google Colab \u7b14\u8bb0\u672c<\/a><\/strong>&nbsp;&#8211; \u8be5\u7b14\u8bb0\u672c\u5c55\u793a\u4e86\u5982\u4f55\u5bf9 4 \u6bd4\u7279\u6a21\u578b\u8fdb\u884c\u63a8\u7406\uff0c\u4ee5\u53ca\u5982\u4f55\u5728\u514d\u8d39\u7684 Google Colab \u5b9e\u4f8b\u4e0a\u8fd0\u884c GPT-neo-X \u6a21\u578b (20B) \u3002<\/li>\n\n\n\n<li><strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/colab.research.google.com\/drive\/1VoYNfYDKcKRQRor98Zbf2-9VQTtGJ24k%3Fusp%3Dsharing\" target=\"_blank\" rel=\"noreferrer noopener\">\u5fae\u8c03\u7684 Google Colab \u7b14\u8bb0\u672c<\/a><\/strong>&nbsp;&#8211; \u8be5\u7b14\u8bb0\u672c\u5c55\u793a\u4e86\u5982\u4f55\u4f7f\u7528 Hugging Face \u751f\u6001\u7cfb\u7edf\u5728\u4e0b\u6e38\u4efb\u52a1\u4e0a\u5fae\u8c03 4 \u6bd4\u7279\u6a21\u578b\u3002\u6211\u4eec\u8bc1\u660e\u4e86\u53ef\u4ee5\u5728 Google Colab \u5b9e\u4f8b\u4e0a\u5fae\u8c03 GPT-neo-X 20B\uff01<\/li>\n\n\n\n<li><strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/artidoro\/qlora\" target=\"_blank\" rel=\"noreferrer noopener\">\u7528\u4e8e\u590d\u73b0\u8bba\u6587\u7ed3\u679c\u7684\u539f\u59cb\u4ee3\u7801\u5e93<\/a><\/strong><\/li>\n\n\n\n<li><strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/huggingface.co\/spaces\/uwnlp\/guanaco-playground-tgi\" target=\"_blank\" rel=\"noreferrer noopener\">Guanaco 33B \u7684\u6f14\u793a\u7a7a\u95f4<\/a><\/strong>&nbsp;&#8211; \u4e0b\u6587\u4e2d\u4e5f\u5305\u542b\u4e86\u8fd9\u4e2a\u6f14\u793a\u7a7a\u95f4\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u5f15\u8a00<\/strong><\/h2>\n\n\n\n<p>\u5982\u679c\u4f60\u5bf9\u6a21\u578b\u7cbe\u5ea6\u53ca\u4e00\u4e9b\u5e38\u89c1\u7684\u6570\u636e\u7c7b\u578b (float16\u3001float32\u3001bfloat16\u3001int8) \u5c1a\u4e0d\u719f\u6089\uff0c\u5efa\u8bae\u4f60\u4ed4\u7ec6\u9605\u8bfb&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/huggingface.co\/blog\/zh\/hf-bitsandbytes-integration\" target=\"_blank\" rel=\"noreferrer noopener\">\u6211\u4eec\u7684\u7b2c\u4e00\u7bc7\u535a\u6587<\/a><\/strong>\uff0c\u8fd9\u7bc7\u535a\u6587\u56fe\u6587\u5e76\u8302\u5730\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u76f8\u5173\u6982\u5ff5\u7684\u7ec6\u8282\u3002<\/p>\n\n\n\n<p>\u5982\u9700\u66f4\u591a\u4fe1\u606f\uff0c\u5efa\u8bae\u67e5\u9605&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/en.wikibooks.org\/wiki\/A-level_Computing\/AQA\/Paper_2\/Fundamentals_of_data_representation\/Floating_point_numbers%23%3A~%3Atext%3DIn%2520decimal%252C%2520very%2520large%2520numbers%2Cbe%2520used%2520for%2520binary%2520numbers.\" target=\"_blank\" rel=\"noreferrer noopener\">\u8fd9\u7bc7 wikibook \u6587\u6863<\/a><\/strong>&nbsp;\u4ee5\u4e86\u89e3\u6d6e\u70b9\u8868\u793a\u7684\u57fa\u7840\u77e5\u8bc6\u3002<\/p>\n\n\n\n<p>QLoRA \u8bba\u6587\u4e2d\u63a2\u8ba8\u4e86\u4e24\u79cd\u4e0d\u540c\u7684\u6570\u636e\u7c7b\u578b: 4 \u6bd4\u7279 Float \u548c 4 \u6bd4\u7279 NormalFloat\u3002\u8fd9\u91cc\u6211\u4eec\u5c06\u8ba8\u8bba 4 \u6bd4\u7279 Float \u6570\u636e\u7c7b\u578b\uff0c\u56e0\u4e3a\u5b83\u66f4\u5bb9\u6613\u7406\u89e3\u3002<\/p>\n\n\n\n<p>FP8 \u548c FP4 \u5206\u522b\u4ee3\u8868\u6d6e\u70b9 8 \u6bd4\u7279\u548c 4 \u6bd4\u7279\u7cbe\u5ea6\u3002\u5b83\u4eec\u5c5e\u4e8e minifloats \u6d6e\u70b9\u503c\u7cfb\u5217 (minifloats \u7cfb\u5217\u8fd8\u5305\u62ec\u5176\u4ed6\u7cbe\u5ea6\uff0c\u5982 bfloat16 \u548c float16)\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u5148\u770b\u4e00\u4e0b\u5982\u4f55\u7528 FP8 \u683c\u5f0f\u8868\u793a\u6d6e\u70b9\u503c\uff0c\u7136\u540e\u4e86\u89e3 FP4 \u683c\u5f0f\u662f\u4ec0\u4e48\u6837\u5b50\u7684\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>FP8 \u683c\u5f0f<\/strong><\/h3>\n\n\n\n<p>\u6b63\u5982\u4e4b\u524d\u7684\u535a\u6587\u4e2d\u6240\u8ba8\u8bba\u7684\uff0cn \u6bd4\u7279\u7684\u6d6e\u70b9\u6570\u4e2d\u6bcf\u4e2a\u6bd4\u7279\u90fd\u5c5e\u4e8e\u4e00\u4e2a\u7279\u5b9a\u7c7b\u522b\uff0c\u8d1f\u8d23\u8868\u793a\u6570\u5b57\u7684\u5404\u4e2a\u7ec4\u6210\u90e8\u5206 (\u7b26\u53f7\u3001\u5c3e\u6570\u548c\u6307\u6570)\u3002<\/p>\n\n\n\n<p><strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/arxiv.org\/pdf\/2209.05433.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">FP8 for Deep Learning<\/a><\/strong>&nbsp;\u8fd9\u7bc7\u8bba\u6587\u9996\u6b21\u5f15\u5165\u4e86 FP8 (\u6d6e\u70b9 8) \u683c\u5f0f\uff0c\u5176\u6709\u4e24\u79cd\u4e0d\u540c\u7684\u7f16\u7801\u65b9\u5f0f: E4M3 (4 \u4f4d\u6307\u6570\uff0c3 \u4f4d\u5c3e\u6570) \u548c E5M2 (5 \u4f4d\u6307\u6570\uff0c2 \u4f4d\u5c3e\u6570)\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1212\" height=\"684\" src=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-2.png\" alt=\"\" class=\"wp-image-2539\" srcset=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-2.png 1212w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-2-300x169.png 300w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-2-1024x578.png 1024w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-2-768x433.png 768w\" sizes=\"auto, (max-width: 1212px) 100vw, 1212px\" \/><figcaption class=\"wp-element-caption\">fp8 \u7f16\u7801\u65b9\u6848 FP8 \u683c\u5f0f\u6982\u89c8\u3002\u56fe\u6e90: \u5185\u5bb9\u6765\u81ea\u4e8e sgugger<\/figcaption><\/figure>\n\n\n\n<p>\u5c3d\u7ba1\u968f\u7740\u6bd4\u7279\u6570\u4ece 32 \u51cf\u5c11\u5230 8\uff0c\u7cbe\u5ea6\u5927\u5927\u964d\u4f4e\u4e86\uff0c\u4f46\u8fd9\u4e24\u79cd 8 \u6bd4\u7279\u7f16\u7801\u4ecd\u6709\u5f88\u591a\u7528\u6b66\u4e4b\u5730\u3002\u76ee\u524d\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/NVIDIA\/TransformerEngine\" target=\"_blank\" rel=\"noreferrer noopener\">Transformer Engine \u5e93<\/a><\/strong>&nbsp;\u6765\u4f7f\u7528\u5b83\u4eec\uff0cHF \u751f\u6001\u7cfb\u7edf\u7684 accelerate \u5e93\u4e5f\u96c6\u6210\u4e86 Transformer Engine \u5e93\u3002<\/p>\n\n\n\n<p>E4M3 \u683c\u5f0f\u53ef\u4ee5\u8868\u793a\u7684\u6d6e\u70b9\u6570\u8303\u56f4\u4e3a -448 \u5230 448\u3002\u800c E5M2 \u683c\u5f0f\u56e0\u4e3a\u589e\u52a0\u4e86\u6307\u6570\u4f4d\u6570\uff0c\u5176\u8868\u793a\u8303\u56f4\u6269\u5927\u4e3a -57344 \u5230 57344 &#8211; \u4f46\u5176\u76f8\u5bf9 E4M3 \u800c\u8a00\u7cbe\u5ea6\u4f1a\u6709\u635f\u5931\uff0c\u56e0\u4e3a\u4e24\u8005\u53ef\u8868\u793a\u7684\u6570\u7684\u4e2a\u6570\u4fdd\u6301\u4e0d\u53d8\u3002\u7ecf\u9a8c\u8bc1\u660e\uff0cE4M3 \u6700\u9002\u5408\u524d\u5411\u8ba1\u7b97\uff0cE5M2 \u6700\u9002\u5408\u540e\u5411\u8ba1\u7b97\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>FP4 \u7cbe\u5ea6\u7b80\u8ff0<\/strong><\/h3>\n\n\n\n<p>\u7b26\u53f7\u4f4d\u8868\u793a\u7b26\u53f7 (+\/-)\uff0c\u6307\u6570\u4f4d\u8f6c\u8bd1\u6210\u8be5\u90e8\u5206\u6240\u8868\u793a\u7684\u6574\u6570\u7684 2 \u6b21\u65b9 (\u4f8b\u5982&nbsp;<code>2^{010} = 2^{2} = 4<\/code>&nbsp;)\u3002\u5206\u6570\u6216\u5c3e\u6570\u4f4d\u8868\u793a\u6210 -2 \u7684\u5e42\u7684\u603b\u548c\uff0c\u5982\u679c\u7b2c i \u4f4d\u4e3a&nbsp;<code>1<\/code>&nbsp;\uff0c\u5219\u548c\u52a0\u4e0a&nbsp;<code>2^-i<\/code>&nbsp;\uff0c\u5426\u5219\u4fdd\u6301\u4e0d\u53d8\uff0c\u8fd9\u91cc i \u662f\u8be5\u4f4d\u5728\u6bd4\u7279\u5e8f\u5217\u4e2d\u7684\u4f4d\u7f6e\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e\u5c3e\u6570 1010\uff0c\u6211\u4eec\u6709&nbsp;<code>(2^-1 + 0 + 2^-3 + 0) = (0.5 + 0.125) = 0.625<\/code>&nbsp;\uff0c\u7136\u540e\uff0c\u6211\u4eec\u7ed9\u5206\u6570\u52a0\u4e0a\u4e00\u4e2a&nbsp;<em>1<\/em>&nbsp;\uff0c\u5f97\u5230&nbsp;<code>1.625<\/code>&nbsp;\u3002\u6700\u540e\uff0c\u518d\u5c06\u6240\u6709\u7ed3\u679c\u76f8\u4e58\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u4f7f\u7528 2 \u4e2a\u6307\u6570\u4f4d\u548c 1 \u4e2a\u5c3e\u6570\u4f4d\uff0c\u7f16\u7801 1101 \u5bf9\u5e94\u7684\u6570\u503c\u4e3a:<\/p>\n\n\n\n<p><code>-1 * 2^(2)*(1 + 2^-1) = -1 * 4 * 1.5 = -6<\/code><\/p>\n\n\n\n<p>FP4 \u6ca1\u6709\u56fa\u5b9a\u7684\u683c\u5f0f\uff0c\u56e0\u6b64\u53ef\u4ee5\u5c1d\u8bd5\u4e0d\u540c\u5c3e\u6570\/\u6307\u6570\u7684\u7ec4\u5408\u3002\u4e00\u822c\u6765\u8bf4\uff0c\u5728\u5927\u591a\u6570\u60c5\u51b5\u4e0b\uff0c3 \u4e2a\u6307\u6570\u4f4d\u7684\u6548\u679c\u4f1a\u66f4\u597d\u4e00\u4e9b\u3002\u4f46\u67d0\u4e9b\u60c5\u51b5\u4e0b\uff0c2 \u4e2a\u6307\u6570\u4f4d\u52a0\u4e0a 1 \u4e2a\u5c3e\u6570\u4f4d\u6027\u80fd\u4f1a\u66f4\u597d\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>QLoRA\uff0c\u7ecf\u7531\u91cf\u5316\u5b9e\u73b0\u5927\u6a21\u578b\u81ea\u7531\u7684\u65b0\u9014\u5f84<\/strong><\/h2>\n\n\n\n<p>\u7b80\u800c\u8a00\u4e4b\uff0c\u4e0e\u6807\u51c6 16 \u6bd4\u7279\u6a21\u578b\u5fae\u8c03\u76f8\u6bd4\uff0cQLoRA \u5728\u4e0d\u727a\u7272\u6027\u80fd\u7684\u524d\u63d0\u4e0b\u51cf\u5c11\u4e86 LLM \u5fae\u8c03\u7684\u5185\u5b58\u4f7f\u7528\u91cf\u3002\u4f7f\u7528\u8be5\u65b9\u6cd5\uff0c\u6211\u4eec\u53ef\u5728\u5355\u4e2a 24GB GPU \u4e0a\u5fae\u8c03 33B \u6a21\u578b\uff0c\u8fd8\u53ef\u4ee5\u5728\u5355\u4e2a 46GB GPU \u4e0a\u5fae\u8c03 65B \u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u66f4\u5177\u4f53\u5730\u8bf4\uff0cQLoRA \u4f7f\u7528 4 \u6bd4\u7279\u91cf\u5316\u6765\u538b\u7f29\u9884\u8bad\u7ec3\u7684\u8bed\u8a00\u6a21\u578b\u3002\u7136\u540e\u51bb\u7ed3\u57fa\u7840\u6a21\u578b\u7684\u53c2\u6570\uff0c\u5e76\u5c06\u76f8\u5bf9\u5c11\u91cf\u7684\u53ef\u8bad\u7ec3\u53c2\u6570\u4ee5\u4f4e\u79e9\u9002\u914d\u5668\u7684\u5f62\u5f0f\u6dfb\u52a0\u5230\u6a21\u578b\u4e2d\u3002\u5728\u5fae\u8c03\u8fc7\u7a0b\u4e2d\uff0cQLoRA \u901a\u8fc7\u51bb\u7ed3\u7684 4 \u6bd4\u7279\u91cf\u5316\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u5c06\u68af\u5ea6\u53cd\u5411\u4f20\u64ad\u5230\u4f4e\u79e9\u9002\u914d\u5668\u4e2d\u3002LoRA \u5c42\u7684\u6743\u91cd\u662f\u8bad\u7ec3\u671f\u95f4\u552f\u4e00\u53ef\u66f4\u65b0\u7684\u53c2\u6570\u3002\u4f60\u53ef\u9605\u8bfb&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/arxiv.org\/abs\/2106.09685\" target=\"_blank\" rel=\"noreferrer noopener\">\u539f\u59cb LoRA \u8bba\u6587<\/a><\/strong>&nbsp;\u4ee5\u4e86\u89e3\u66f4\u591a\u6709\u5173 LoRA \u7684\u4fe1\u606f\u3002<\/p>\n\n\n\n<p>QLoRA \u6709\u4e00\u4e2a\u7528\u4e8e\u5b58\u50a8\u57fa\u7840\u6a21\u578b\u6743\u91cd\u7684\u6570\u636e\u7c7b\u578b (\u901a\u5e38\u4e3a 4 \u6bd4\u7279 NormalFloat) \u548c\u4e00\u4e2a\u7528\u4e8e\u6267\u884c\u8ba1\u7b97\u7684\u6570\u636e\u7c7b\u578b (16 \u6bd4\u7279 BrainFloat)\u3002QLoRA \u5c06\u6743\u91cd\u4ece\u5b58\u50a8\u6570\u636e\u7c7b\u578b\u53cd\u91cf\u5316\u4e3a\u8ba1\u7b97\u6570\u636e\u7c7b\u578b\uff0c\u4ee5\u6267\u884c\u524d\u5411\u548c\u540e\u5411\u4f20\u64ad\uff0c\u4f46\u4ec5\u8ba1\u7b97 bfloat16 \u7684 LoRA \u53c2\u6570\u7684\u6743\u91cd\u68af\u5ea6\u3002\u6743\u91cd\u4ec5\u5728\u9700\u8981\u65f6\u624d\u89e3\u538b\u7f29\uff0c\u56e0\u6b64\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u671f\u95f4\u5185\u5b58\u4f7f\u7528\u7387\u90fd\u80fd\u4fdd\u6301\u8f83\u4f4e\u6c34\u5e73\u3002<\/p>\n\n\n\n<p>\u5e7f\u6cdb\u7684\u5b9e\u9a8c\u8868\u660e QLoRA \u5fae\u8c03\u4e0e 16 \u6bd4\u7279\u5fae\u8c03\u7684\u6027\u80fd\u65d7\u9f13\u76f8\u5f53\u3002\u6b64\u5916\uff0c\u5728&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/huggingface.co\/datasets\/OpenAssistant\/oasst1\" target=\"_blank\" rel=\"noreferrer noopener\">OpenAssistant \u6570\u636e\u96c6 (OASST1)<\/a><\/strong>&nbsp;\u4e0a\u5bf9 LLaMA \u6a21\u578b\u4f7f\u7528 QLoRA \u5fae\u8c03\u800c\u5f97\u7684 Guanaco \u6a21\u578b\u662f\u76ee\u524d\u6700\u5148\u8fdb\u7684\u804a\u5929\u673a\u5668\u4eba\u7cfb\u7edf\uff0c\u5176\u5728 Vicuna \u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u63a5\u8fd1 ChatGPT\u3002\u8fd9\u662f QLoRA \u5fae\u8c03\u5a01\u529b\u7684\u8fdb\u4e00\u6b65\u5c55\u793a\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u5982\u4f55\u5728 transformers \u4e2d\u4f7f\u7528\u5b83\uff1f<\/strong><\/h2>\n\n\n\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u4ecb\u7ecd\u8be5\u65b9\u6cd5\u5728&nbsp;<code>transformers<\/code>&nbsp;\u4e2d\u7684\u96c6\u6210\u3001\u5982\u4f55\u4f7f\u7528\u5b83\u4ee5\u53ca\u76ee\u524d\u652f\u6301\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u5165\u95e8<\/strong><\/h3>\n\n\n\n<p>\u4f5c\u4e3a\u5feb\u901f\u5165\u95e8\uff0c\u6211\u4eec\u53ef\u4ee5\u4ece\u6e90\u4ee3\u7801\u5b89\u88c5&nbsp;<code>accelerate<\/code>&nbsp;\u548c&nbsp;<code>transformers<\/code>&nbsp;\uff0c\u4ee5\u52a0\u8f7d 4 \u6bd4\u7279\u6a21\u578b\uff0c\u53e6\u8bf7\u786e\u4fdd\u5df2\u5b89\u88c5\u6700\u65b0\u7248\u672c\u7684&nbsp;<code>bitsandbytes<\/code>&nbsp;\u5e93 (0.39.0)\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>pip install -q -U bitsandbytes\npip install -q -U git+https:\/\/github.com\/huggingface\/transformers.git\npip install -q -U git+https:\/\/github.com\/huggingface\/peft.git\npip install -q -U git+https:\/\/github.com\/huggingface\/accelerate.git<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u5feb\u901f\u5f00\u59cb<\/strong><\/h3>\n\n\n\n<p>\u4ee5 4 \u6bd4\u7279\u52a0\u8f7d\u6a21\u578b\u7684\u57fa\u672c\u65b9\u6cd5\u662f\u901a\u8fc7\u5728\u8c03\u7528&nbsp;<code>from_pretrained<\/code>&nbsp;\u65b9\u6cd5\u65f6\u4f20\u9012\u53c2\u6570&nbsp;<code>load_in_4bit=True<\/code>&nbsp;\uff0c\u5e76\u5c06\u8bbe\u5907\u6620\u5c04\u8bbe\u7f6e\u6210&nbsp;<code>\u201cauto\u201d<\/code>&nbsp;\u3002<\/p>\n\n\n\n<div class=\"wp-block-urvanov-syntax-highlighter-code-block\"><pre class=\"lang:python decode:true \" >from transformers import AutoModelForCausalLM\n\nmodel = AutoModelForCausalLM.from_pretrained(\"facebook\/opt-350m\", load_in_4bit=True, device_map=\"auto\")\n...<\/pre><\/div>\n\n\n\n<p>\u8fd9\u6837\u5c31\u884c\u4e86\uff01<\/p>\n\n\n\n<p>\u4e00\u822c\u5730\uff0c\u6211\u4eec\u5efa\u8bae\u7528\u6237\u5728\u4f7f\u7528&nbsp;<code>device_map<\/code>&nbsp;\u52a0\u8f7d\u6a21\u578b\u540e\u4e0d\u8981\u624b\u52a8\u8bbe\u7f6e\u8bbe\u5907\u3002\u56e0\u6b64\uff0c\u5728\u8be5\u884c\u4e4b\u540e\u5e94\u907f\u514d\u5bf9\u6a21\u578b\u6216\u6a21\u578b\u7684\u4efb\u4f55\u5b50\u6a21\u5757\u8fdb\u884c\u4efb\u4f55\u8bbe\u5907\u5206\u914d &#8211; \u9664\u975e\u4f60\u77e5\u9053\u81ea\u5df1\u5728\u505a\u4ec0\u4e48\u3002<\/p>\n\n\n\n<p>\u8bf7\u8bb0\u4f4f\uff0c\u52a0\u8f7d\u91cf\u5316\u6a21\u578b\u4f1a\u81ea\u52a8\u5c06\u6a21\u578b\u7684\u5176\u4ed6\u5b50\u6a21\u5757\u8f6c\u6362\u4e3a&nbsp;<code>float16<\/code>&nbsp;\u6570\u636e\u7c7b\u578b\u3002\u4f60\u53ef\u4ee5\u901a\u8fc7\u5c06&nbsp;<code>torch_dtype=dtype<\/code>&nbsp;\u4f20\u9012\u7ed9&nbsp;<code>from_pretrained<\/code>&nbsp;\u65b9\u6cd5\u6765\u4fee\u6539\u6b64\u884c\u4e3a (\u4f8b\u5982\uff0c\u5982\u679c\u4f60\u5e0c\u671b\u5728\u5c42\u89c4\u4e00\u5316\u7b97\u5b50\u4e2d\u4f7f\u7528&nbsp;<code>float32<\/code>&nbsp;)\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u9ad8\u7ea7\u7528\u6cd5<\/strong><\/h3>\n\n\n\n<p>\u4f60\u53ef\u4ee5\u4f7f\u7528 4 \u6bd4\u7279\u91cf\u5316\u7684\u4e0d\u540c\u53d8\u4f53\uff0c\u4f8b\u5982 NF4 (NormalFloat4 (\u9ed8\u8ba4) ) \u6216\u7eaf FP4 \u91cf\u5316\u3002\u4ece\u7406\u8bba\u5206\u6790\u548c\u5b9e\u8bc1\u7ed3\u679c\u6765\u770b\uff0c\u6211\u4eec\u5efa\u8bae\u4f7f\u7528 NF4 \u91cf\u5316\u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<p>\u5176\u4ed6\u9009\u9879\u5305\u62ec&nbsp;<code>bnb_4bit_use_double_quant<\/code>&nbsp;\uff0c\u5b83\u5728\u7b2c\u4e00\u8f6e\u91cf\u5316\u4e4b\u540e\u4f1a\u8fdb\u884c\u7b2c\u4e8c\u8f6e\u91cf\u5316\uff0c\u4e3a\u6bcf\u4e2a\u53c2\u6570\u989d\u5916\u8282\u7701 0.4 \u6bd4\u7279\u3002\u6700\u540e\u662f\u8ba1\u7b97\u7c7b\u578b\uff0c\u867d\u7136 4 \u6bd4\u7279 bitsandbytes \u4ee5 4 \u6bd4\u7279\u5b58\u50a8\u6743\u91cd\uff0c\u4f46\u8ba1\u7b97\u4ecd\u7136\u4ee5 16 \u6216 32 \u6bd4\u7279\u8fdb\u884c\uff0c\u8fd9\u91cc\u53ef\u4ee5\u9009\u62e9\u4efb\u610f\u7ec4\u5408 (float16\u3001bfloat16\u3001float32 \u7b49)\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u4f7f\u7528 16 \u6bd4\u7279\u8ba1\u7b97\u6570\u636e\u7c7b\u578b (\u9ed8\u8ba4 torch.float32)\uff0c\u77e9\u9635\u4e58\u6cd5\u548c\u8bad\u7ec3\u5c06\u4f1a\u66f4\u5feb\u3002\u7528\u6237\u5e94\u8be5\u5229\u7528 transformers \u4e2d\u6700\u65b0\u7684&nbsp;<code>BitsAndBytesConfig<\/code>&nbsp;\u6765\u66f4\u6539\u8fd9\u4e9b\u53c2\u6570\u3002\u4e0b\u9762\u662f\u4f7f\u7528 NF4 \u91cf\u5316\u52a0\u8f7d 4 \u6bd4\u7279\u6a21\u578b\u7684\u793a\u4f8b\uff0c\u4f8b\u5b50\u4e2d\u4f7f\u7528\u4e86\u53cc\u91cf\u5316\u4ee5\u53ca bfloat16 \u8ba1\u7b97\u6570\u636e\u7c7b\u578b\u4ee5\u52a0\u901f\u8bad\u7ec3:<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>from transformers import BitsAndBytesConfig\n\nnf4_config = BitsAndBytesConfig(\n   load_in_4bit=True,\n   bnb_4bit_quant_type=\"nf4\",\n   bnb_4bit_use_double_quant=True,\n   bnb_4bit_compute_dtype=torch.bfloat16\n)\n\nmodel_nf4 = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=nf4_config)<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u66f4\u6539\u8ba1\u7b97\u6570\u636e\u7c7b\u578b<\/strong><\/h3>\n\n\n\n<p>\u5982\u4e0a\u6240\u8ff0\uff0c\u4f60\u8fd8\u53ef\u4ee5\u901a\u8fc7\u66f4\u6539&nbsp;<code>BitsAndBytesConfig<\/code>&nbsp;\u4e2d\u7684&nbsp;<code>bnb_4bit_compute_dtype<\/code>&nbsp;\u53c2\u6570\u6765\u66f4\u6539\u91cf\u5316\u6a21\u578b\u7684\u8ba1\u7b97\u6570\u636e\u7c7b\u578b\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import torch\nfrom transformers import BitsAndBytesConfig\n\nquantization_config = BitsAndBytesConfig(\n   load_in_4bit=True,\n   bnb_4bit_compute_dtype=torch.bfloat16\n)<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u5d4c\u5957\u91cf\u5316<\/strong><\/h3>\n\n\n\n<p>\u8981\u542f\u7528\u5d4c\u5957\u91cf\u5316\uff0c\u4f60\u53ef\u4ee5\u4f7f\u7528&nbsp;<code>BitsAndBytesConfig<\/code>&nbsp;\u4e2d\u7684&nbsp;<code>bnb_4bit_use_double_quant<\/code>&nbsp;\u53c2\u6570\u3002\u8fd9\u5c06\u4f1a\u5728\u7b2c\u4e00\u8f6e\u91cf\u5316\u4e4b\u540e\u542f\u7528\u7b2c\u4e8c\u8f6e\u91cf\u5316\uff0c\u4ee5\u4fbf\u6bcf\u4e2a\u53c2\u6570\u989d\u5916\u8282\u7701 0.4 \u6bd4\u7279\u3002\u6211\u4eec\u5728\u4e0a\u6587\u63d0\u53ca\u7684\u5fae\u8c03 Google Colab \u7b14\u8bb0\u672c\u4e2d\u4e5f\u4f7f\u7528\u4e86\u6b64\u529f\u80fd\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>from transformers import BitsAndBytesConfig\n\ndouble_quant_config = BitsAndBytesConfig(\n   load_in_4bit=True,\n   bnb_4bit_use_double_quant=True,\n)\n\nmodel_double_quant = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=double_quant_config)<\/code><\/pre>\n\n\n\n<p>\u5f53\u7136\uff0c\u6b63\u5982\u672c\u8282\u5f00\u5934\u63d0\u5230\u7684\uff0c\u6240\u6709\u8fd9\u4e9b\u529f\u80fd\u90fd\u662f\u53ef\u4efb\u610f\u7ec4\u5408\u7684\u3002\u4f60\u53ef\u4ee5\u5c06\u6240\u6709\u8fd9\u4e9b\u53c2\u6570\u7ec4\u5408\u5728\u4e00\u8d77\uff0c\u627e\u5230\u6700\u9002\u5408\u4f60\u7684\u914d\u7f6e\u3002\u7ecf\u9a8c\u6cd5\u5219\u662f: \u5982\u679c\u5185\u5b58\u6709\u9650\u5236\uff0c\u4f7f\u7528\u53cc\u91cf\u5316; \u4f7f\u7528 NF4 \u4ee5\u83b7\u5f97\u66f4\u9ad8\u7684\u7cbe\u5ea6; \u4f7f\u7528 16 \u6bd4\u7279\u6d6e\u70b9\u52a0\u5feb\u5fae\u8c03\u901f\u5ea6\u3002\u4f5c\u4e3a\u4e00\u4e2a\u4f8b\u5b50\uff0c\u5728&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/colab.research.google.com\/drive\/1ge2F1QSK8Q7h0hn3YKuBCOAS0bK8E0wf%3Fusp%3Dsharing\" target=\"_blank\" rel=\"noreferrer noopener\">\u63a8\u7406\u6f14\u793a\u5e94\u7528<\/a><\/strong>&nbsp;\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528\u5d4c\u5957\u91cf\u5316\u3001bfloat16 \u8ba1\u7b97\u6570\u636e\u7c7b\u578b\u4ee5\u53ca NF4 \u91cf\u5316\u5728\u5355\u4e2a 16GB GPU \u4e2d\u4f7f\u7528 4 \u6bd4\u7279\u5b8c\u6210\u4e86\u5bf9 gpt-neo-x-20b (40GB) \u6a21\u578b\u7684\u62df\u5408\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u5e38\u89c1\u95ee\u9898<\/strong><\/h3>\n\n\n\n<p>\u672c\u8282\u6211\u4eec\u6765\u56de\u7b54\u4e00\u4e9b\u5e38\u89c1\u95ee\u9898\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>FP4 \u91cf\u5316\u6709\u786c\u4ef6\u8981\u6c42\u5417\uff1f<\/strong><\/h3>\n\n\n\n<p>\u8bf7\u6ce8\u610f\uff0c\u6b64\u65b9\u6cd5\u4ec5\u4e0e GPU \u517c\u5bb9\uff0c\u76ee\u524d\u5c1a\u65e0\u6cd5\u5728 CPU \u4e0a\u5bf9\u6a21\u578b\u8fdb\u884c 4 \u6bd4\u7279\u91cf\u5316\u3002\u5728 GPU \u4e2d\uff0c\u6b64\u65b9\u6cd5\u6ca1\u6709\u4efb\u4f55\u786c\u4ef6\u8981\u6c42\uff0c\u53ea\u8981\u5b89\u88c5\u4e86 CUDA&gt;=11.2\uff0c\u4efb\u4f55 GPU \u90fd\u53ef\u4ee5\u7528\u4e8e\u8fd0\u884c 4 \u6bd4\u7279\u91cf\u5316\u3002<\/p>\n\n\n\n<p>\u53e6\u8bf7\u8bb0\u4f4f\uff0c\u8ba1\u7b97\u4e0d\u662f\u4ee5 4 \u6bd4\u7279\u5b8c\u6210\u7684\uff0c\u4ec5\u4ec5\u662f\u6743\u91cd\u548c\u6fc0\u6d3b\u88ab\u538b\u7f29\u4e3a\u8be5\u683c\u5f0f\uff0c\u800c\u8ba1\u7b97\u4ecd\u5728\u6307\u5b9a\u7684\u6216\u8005\u539f\u59cb\u6570\u636e\u7c7b\u578b\u4e0a\u8fdb\u884c\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u652f\u6301\u54ea\u4e9b\u6a21\u578b\uff1f<\/strong><\/h3>\n\n\n\n<p>\u4e0e&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/huggingface.co\/blog\/zh\/hf-bitsandbytes-integration\" target=\"_blank\" rel=\"noreferrer noopener\">\u8fd9\u7bc7\u535a\u6587<\/a><\/strong>&nbsp;\u4e2d\u4ecb\u7ecd\u7684 LLM.int8 \u96c6\u6210\u7c7b\u4f3c\uff0c\u6211\u4eec\u7684\u96c6\u6210\u4e25\u91cd\u4f9d\u8d56\u4e8e&nbsp;<code>accelerate<\/code>&nbsp;\u5e93\u3002\u56e0\u6b64\uff0c\u4efb\u4f55\u652f\u6301&nbsp;<code>accelerate<\/code>&nbsp;\u5e93\u52a0\u8f7d\u7684\u6a21\u578b (\u5373\u8c03\u7528&nbsp;<code>from_pretrained<\/code>&nbsp;\u65f6\u652f\u6301&nbsp;<code>device_map<\/code>&nbsp;\u53c2\u6570) \u90fd\u53ef\u4ee5\u8fdb\u884c 4 \u6bd4\u7279\u91cf\u5316\u3002\u53e6\u8bf7\u6ce8\u610f\uff0c\u8fd9\u4e0e\u6a21\u6001\u5b8c\u5168\u65e0\u5173\uff0c\u53ea\u8981\u53ef\u4ee5\u4f7f\u7528&nbsp;<code>device_map<\/code>&nbsp;\u53c2\u6570\u52a0\u8f7d\u6a21\u578b\uff0c\u5c31\u53ef\u4ee5\u91cf\u5316\u5b83\u4eec\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u6587\u672c\u6a21\u578b\uff0c\u622a\u81f3\u672c\u6587\u64b0\u5199\u65f6\uff0c\u6700\u5e38\u7528\u7684\u67b6\u6784\u90fd\u662f\u652f\u6301\u7684\uff0c\u4f8b\u5982\u7528\u4e8e\u7eaf\u6587\u672c\u7684 Llama\u3001OPT\u3001GPT-Neo\u3001GPT-NeoX\uff0c\u7528\u4e8e\u591a\u6a21\u6001\u7684 Blip2 \u7b49\u3002<\/p>\n\n\n\n<p>\u622a\u81f3\u672c\u6587\u64b0\u5199\u65f6\uff0c\u652f\u6301&nbsp;<code>accelerate<\/code>&nbsp;\u7684\u6a21\u578b\u6709:<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>&#91;\n    'bigbird_pegasus', 'blip_2', 'bloom', 'bridgetower', 'codegen', 'deit', 'esm',\n    'gpt2', 'gpt_bigcode', 'gpt_neo', 'gpt_neox', 'gpt_neox_japanese', 'gptj', 'gptsan_japanese',\n    'lilt', 'llama', 'longformer', 'longt5', 'luke', 'm2m_100', 'mbart', 'mega', 'mt5', 'nllb_moe',\n    'open_llama', 'opt', 'owlvit', 'plbart', 'roberta', 'roberta_prelayernorm', 'rwkv', 'switch_transformers',\n    't5', 'vilt', 'vit', 'vit_hybrid', 'whisper', 'xglm', 'xlm_roberta'\n]<\/code><\/pre>\n\n\n\n<p>\u8bf7\u6ce8\u610f\uff0c\u5982\u679c\u4f60\u6700\u559c\u6b22\u7684\u6a21\u578b\u4e0d\u5728\u5217\u8868\u4e2d\uff0c\u4f60\u53ef\u4ee5\u63d0\u4ea4\u4e00\u4e2a PR \u6216\u5728 transformers \u4e2d\u63d0\u4ea4\u4e00\u4e2a\u95ee\u9898\uff0c\u4ee5\u6dfb\u52a0\u5bf9\u8be5\u67b6\u6784\u7684 accelerate \u52a0\u8f7d\u7684\u652f\u6301\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u6211\u4eec\u53ef\u4ee5\u8bad\u7ec3 4 \u6bd4\u7279 \/ 8 \u6bd4\u7279\u6a21\u578b\u5417\uff1f<\/strong><\/h3>\n\n\n\n<p>\u5bf9\u8fd9\u4e9b\u6a21\u578b\u8fdb\u884c\u5168\u6a21\u578b 4 \u6bd4\u7279\u8bad\u7ec3\u662f\u4e0d\u53ef\u80fd\u7684\u3002\u4f46\u662f\uff0c\u4f60\u53ef\u4ee5\u5229\u7528\u53c2\u6570\u9ad8\u6548\u5fae\u8c03 (PEFT) \u6765\u8bad\u7ec3\u8fd9\u4e9b\u6a21\u578b\uff0c\u5373\u5728\u57fa\u7840\u6a21\u578b\u4e4b\u4e0a\u8bad\u7ec3\u65b0\u589e\u90e8\u5206\u5982\u9002\u914d\u5668\u3002QLoRA \u8bba\u6587\u5c31\u662f\u8fd9\u4e48\u505a\u7684\uff0cHugging Face \u7684 PEFT \u5e93\u4e5f\u6b63\u5f0f\u652f\u6301\u4e86\u8be5\u65b9\u6cd5\u3002\u6211\u4eec\u63d0\u4f9b\u4e86\u76f8\u5e94\u7684&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/colab.research.google.com\/drive\/1VoYNfYDKcKRQRor98Zbf2-9VQTtGJ24k%3Fusp%3Dsharing\" target=\"_blank\" rel=\"noreferrer noopener\">\u5fae\u8c03\u7b14\u8bb0\u672c<\/a><\/strong>\u3002\u5982\u679c\u5927\u5bb6\u60f3\u8981\u590d\u73b0\u8bba\u6587\u7684\u7ed3\u679c\uff0c\u8fd8\u53ef\u4ee5\u67e5\u9605&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/github.com\/artidoro\/qlora\" target=\"_blank\" rel=\"noreferrer noopener\">QLoRA \u4ee3\u7801\u5e93<\/a><\/strong>\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"600\" height=\"439\" src=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image.jpeg\" alt=\"\u52a8\u56fe\u5c01\u9762\" class=\"wp-image-2537\" srcset=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image.jpeg 600w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-300x220.jpeg 300w\" sizes=\"auto, (max-width: 600px) 100vw, 600px\" \/><figcaption class=\"wp-element-caption\">lora \u56fe\u4f8b \u539f\u59cb (\u51bb\u7ed3\u7684) \u9884\u8bad\u7ec3\u6743\u91cd (\u5de6\u4fa7) \u7684\u8f93\u51fa\u6fc0\u6d3b\u9700\u8981\u52a0\u4e0a\u4f4e\u79e9\u9002\u914d\u5668\u7684\u8f93\u51fa\uff0c\u8fd9\u4e2a\u4f4e\u79e9\u9002\u914d\u5668\u7531\u77e9\u9635 A \u548c B \u6743\u91cd\u7ec4\u6210 (\u53f3\u4fa7)\u3002<\/figcaption><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>\u8fd9\u9879\u5de5\u4f5c\u8fd8\u6709\u4ec0\u4e48\u5176\u4ed6\u610f\u4e49\uff1f<\/strong><\/h3>\n\n\n\n<p>\u8fd9\u9879\u5de5\u4f5c\u53ef\u4ee5\u4e3a\u793e\u533a\u548c\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u5e26\u6765\u4e00\u4e9b\u79ef\u6781\u7684\u5f71\u54cd\uff0c\u56e0\u4e3a\u5b83\u53ef\u4ee5\u5f71\u54cd\u5f88\u591a\u53ef\u80fd\u7684\u7528\u6cd5\u6216\u5e94\u7528\u573a\u666f\u3002\u5728 RLHF (\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60) \u4e2d\uff0c\u53ef\u4ee5\u52a0\u8f7d\u5355\u4e2a 4 \u6bd4\u7279\u57fa\u7840\u6a21\u578b\uff0c\u5e76\u5728\u5176\u4e0a\u8bad\u7ec3\u591a\u4e2a\u9002\u914d\u5668\uff0c\u4e00\u4e2a\u7528\u4e8e\u5956\u52b1\u5efa\u6a21\uff0c\u53e6\u4e00\u4e2a\u7528\u4e8e\u4ef7\u503c\u7b56\u7565\u8bad\u7ec3\u3002\u6211\u4eec\u5f88\u5feb\u5c31\u4f1a\u53d1\u5e03\u5173\u4e8e\u6b64\u7528\u6cd5\u7684\u66f4\u8be6\u7ec6\u7684\u535a\u6587\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u8fd8\u9488\u5bf9\u8fd9\u79cd\u91cf\u5316\u65b9\u6cd5\u5bf9\u5728\u6d88\u8d39\u7c7b\u786c\u4ef6\u4e0a\u8bad\u7ec3\u5927\u6a21\u578b\u7684\u5f71\u54cd\u6d89\u53ca\u4e86\u4e00\u4e9b\u57fa\u51c6\u6d4b\u8bd5\u3002\u6211\u4eec\u5728\u82f1\u4f1f\u8fbe T4 (16GB) \u4e0a\u5bf9 2 \u79cd\u4e0d\u540c\u7684\u67b6\u6784 Llama 7B (fp16 \u65f6\uff0c\u6a21\u578b\u5927\u5c0f\u4e3a 15GB) \u548c Llama 13B (fp16 \u65f6\uff0c\u6a21\u578b\u5927\u5c0f\u4e3a 27GB) \u8fdb\u884c\u4e86\u591a\u6b21\u5fae\u8c03\u5b9e\u9a8c\uff0c\u7ed3\u679c\u5982\u4e0b:<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><tbody><tr><th>\u6a21\u578b<\/th><th>\u534a\u7cbe\u5ea6\u6a21\u578b\u5927\u5c0f (GB)<\/th><th>\u786c\u4ef6 \/ \u603b\u663e\u5b58<\/th><th>\u91cf\u5316\u65b9\u6cd5 (CD = \u8ba1\u7b97\u6570\u636e\u7c7b\u578b \/ GC = \u68af\u5ea6 checkpointing \/ NQ = \u53cc\u91cf\u5316)<\/th><th>batch size<\/th><th>\u68af\u5ea6\u7d2f\u79ef\u6b65\u6570<\/th><th>\u4f18\u5316\u5668<\/th><th>\u5e8f\u5217\u957f\u5ea6<\/th><th>\u7ed3\u679c<\/th><\/tr><tr><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>&lt;10B \u6a21\u578b<\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>decapoda-research\/llama-7b-hf<\/td><td>14GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>LLM.int8 (8-bit) + GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>512<\/td><td>\u65e0 OOM<\/td><\/tr><tr><td>decapoda-research\/llama-7b-hf<\/td><td>14GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>LLM.int8 (8-bit) + GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>1024<\/td><td>OOM<\/td><\/tr><tr><td>decapoda-research\/llama-7b-hf<\/td><td>14GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>4bit + NF4 + bf16 CD + no GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>512<\/td><td>\u65e0 OOM<\/td><\/tr><tr><td>decapoda-research\/llama-7b-hf<\/td><td>14GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>4bit + FP4 + bf16 CD + no GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>512<\/td><td>\u65e0 OOM<\/td><\/tr><tr><td>decapoda-research\/llama-7b-hf<\/td><td>14GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>4bit + NF4 + bf16 CD + no GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>1024<\/td><td>OOM<\/td><\/tr><tr><td>decapoda-research\/llama-7b-hf<\/td><td>14GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>4bit + FP4 + bf16 CD + no GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>1024<\/td><td>OOM<\/td><\/tr><tr><td>decapoda-research\/llama-7b-hf<\/td><td>14GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>4bit + NF4 + bf16 CD + GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>1024<\/td><td>\u65e0 OOM<\/td><\/tr><tr><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>10B+ \u6a21\u578b<\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>decapoda-research\/llama-13b-hf<\/td><td>27GB<\/td><td>2xNVIDIA-T4 \/ 32GB<\/td><td>LLM.int8 (8-bit) + GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>512<\/td><td>\u65e0 OOM<\/td><\/tr><tr><td>decapoda-research\/llama-13b-hf<\/td><td>27GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>LLM.int8 (8-bit) + GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>512<\/td><td>OOM<\/td><\/tr><tr><td>decapoda-research\/llama-13b-hf<\/td><td>27GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>4bit + FP4 + bf16 CD + no GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>512<\/td><td>OOM<\/td><\/tr><tr><td>decapoda-research\/llama-13b-hf<\/td><td>27GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>4bit + FP4 + fp16 CD + no GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>512<\/td><td>OOM<\/td><\/tr><tr><td>decapoda-research\/llama-13b-hf<\/td><td>27GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>4bit + NF4 + fp16 CD + GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>512<\/td><td>\u65e0 OOM<\/td><\/tr><tr><td>decapoda-research\/llama-13b-hf<\/td><td>27GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>4bit + NF4 + fp16 CD + GC<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>1024<\/td><td>OOM<\/td><\/tr><tr><td>decapoda-research\/llama-13b-hf<\/td><td>27GB<\/td><td>1xNVIDIA-T4 \/ 16GB<\/td><td>4bit + NF4 + fp16 CD + GC + NQ<\/td><td>1<\/td><td>4<\/td><td>AdamW<\/td><td>1024<\/td><td>\u65e0 OOM<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u6211\u4eec\u4f7f\u7528\u4e86 TRL \u5e93\u4e2d\u6700\u65b0\u7684&nbsp;<code>SFTTrainer<\/code>&nbsp;\uff0c\u4f60\u53ef\u4ee5\u5728&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/gist.github.com\/younesbelkada\/f48af54c74ba6a39a7ae4fd777e72fe8\" target=\"_blank\" rel=\"noreferrer noopener\">\u6b64\u5904<\/a><\/strong>&nbsp;\u627e\u5230\u57fa\u51c6\u6d4b\u8bd5\u811a\u672c\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u6f14\u793a\u7a7a\u95f4<\/strong><\/h2>\n\n\n\n<p>\u60f3\u8bd5\u8bd5\u8bba\u6587\u4e2d\u7684 Guananco \u6a21\u578b\u7684\u8bdd\uff0c\u53ef\u4ee5\u73a9\u73a9\u8fd9\u4e2a&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/huggingface.co\/spaces\/uwnlp\/guanaco-playground-tgi\" target=\"_blank\" rel=\"noreferrer noopener\">\u6f14\u793a\u7a7a\u95f4<\/a><\/strong>\uff0c\u6211\u4eec\u8fd8\u628a\u5b83\u76f4\u63a5\u5d4c\u5165\u5230\u4e86\u4e0b\u9762\u4f9b\u4f60\u76f4\u63a5\u628a\u73a9\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img loading=\"lazy\" decoding=\"async\" width=\"1338\" height=\"1265\" src=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-1.png\" alt=\"\" class=\"wp-image-2538\" srcset=\"https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-1.png 1338w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-1-300x284.png 300w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-1-1024x968.png 1024w, https:\/\/www.aqwu.net\/wp\/wp-content\/uploads\/2024\/03\/image-1-768x726.png 768w\" sizes=\"auto, (max-width: 1338px) 100vw, 1338px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>\u81f4\u8c22<\/strong><\/h2>\n\n\n\n<p>HF \u56e2\u961f\u611f\u8c22\u534e\u76db\u987f\u5927\u5b66\u53c2\u4e0e\u8be5\u9879\u76ee\u7684\u6240\u6709\u4eba\u5458\uff0c\u611f\u8c22\u4ed6\u4eec\u5411\u793e\u533a\u8d21\u732e\u4e86\u4ed6\u4eec\u7684\u5de5\u4f5c\u3002<\/p>\n\n\n\n<p>\u4f5c\u8005\u8fd8\u8981\u611f\u8c22&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/huggingface.co\/pcuenq\" target=\"_blank\" rel=\"noreferrer noopener\">Pedro Cuenca<\/a><\/strong>&nbsp;\u5e2e\u5fd9\u5ba1\u9605\u4e86\u535a\u6587\uff0c\u5e76\u611f\u8c22&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/huggingface.co\/olivierdehaene\" target=\"_blank\" rel=\"noreferrer noopener\">Olivier Dehaene<\/a><\/strong>&nbsp;\u548c&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/huggingface.co\/osanseviero\" target=\"_blank\" rel=\"noreferrer noopener\">Omar Sanseviero<\/a><\/strong>&nbsp;\u5bf9\u5728 HF Hub \u4e0a\u96c6\u6210\u8be5\u8bba\u6587\u63d0\u4f9b\u4e86\u5feb\u901f\u800c\u6709\u529b\u7684\u652f\u6301\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u82f1\u6587\u539f\u6587:&nbsp;<strong><a href=\"https:\/\/link.zhihu.com\/?target=https%3A\/\/hf.co\/blog\/4bit-transformers-bitsandbytes\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/hf.co\/blog\/4bit-transformers-bitsandbytes<\/a><\/strong><br>\u539f\u6587\u4f5c\u8005: Younes Belkada\uff0cTim Dettmers\uff0cArtidoro Pagnoni\uff0cSylvain Gugger\uff0cSourab Mangrulkar<br>\u8bd1\u8005: Matrix Yao (\u59da\u4f1f\u5cf0)\uff0c\u82f1\u7279\u5c14\u6df1\u5ea6\u5b66\u4e60\u5de5\u7a0b\u5e08\uff0c\u5de5\u4f5c\u65b9\u5411\u4e3a transformer-family \u6a21\u578b\u5728\u5404\u6a21\u6001\u6570\u636e\u4e0a\u7684\u5e94\u7528\u53ca\u5927\u89c4\u6a21\u6a21\u578b\u7684\u8bad\u7ec3\u63a8\u7406\u3002<br>\u5ba1\u6821\/\u6392\u7248: zhongdongy (\u963f\u4e1c)<\/p>\n\n\n\n<p>\u53c2\u8003\u94fe\u63a5\uff1a<a href=\"https:\/\/zhuanlan.zhihu.com\/p\/665601576\">\u7528 bitsandbytes\u30014 \u6bd4\u7279\u91cf\u5316\u548c QLoRA \u6253\u9020\u4eb2\u6c11\u7684 LLM &#8211; \u77e5\u4e4e (zhihu.com)<\/a><\/p>\n\n\n\n<p><\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>Hugging Face\u200b \u5df2\u8ba4\u8bc1\u8d26\u53f7 47 \u4eba\u8d5e\u540c\u4e86\u8be5\u6587\u7ae0 \u4f17\u6240\u5468\u77e5\uff0cLLM \u89c4\u6a21\u5e9e\u5927\uff0c\u5982\u679c\u5728\u4e5f\u80fd\u6d88\u8d39\u7c7b\u786c\u4ef6 [&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":"","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":"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":""}},"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[313,289,312],"tags":[393,242,314],"class_list":["post-2535","post","type-post","status-publish","format-standard","hentry","category-chatgpt","category-gpt","category-openai","tag-bitsandbytes","tag-chatgpt","tag-openai-api"],"views":2543,"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2535","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2535"}],"version-history":[{"count":3,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2535\/revisions"}],"predecessor-version":[{"id":2543,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2535\/revisions\/2543"}],"wp:attachment":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2535"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2535"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2535"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}