{"id":4455,"date":"2024-08-15T13:08:26","date_gmt":"2024-08-15T05:08:26","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=4455"},"modified":"2024-08-15T20:00:16","modified_gmt":"2024-08-15T12:00:16","slug":"%e4%bd%bf%e7%94%a8-unsloth-%e8%b6%85%e9%ab%98%e6%95%88%e5%be%ae%e8%b0%83-llama-3-1","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=4455","title":{"rendered":"\u4f7f\u7528 Unsloth \u8d85\u9ad8\u6548\u5fae\u8c03 Llama 3.1"},"content":{"rendered":"\n

\u6700\u5148\u8fdb\u7684\u76d1\u7763\u5fae\u8c03\u521d\u5b66\u8005\u6307\u5357<\/em><\/p>\n\n\n\n

\u6700\u8fd1\u53d1\u5e03\u7684 Llama 3.1 \u4e3a\u6a21\u578b\u63d0\u4f9b\u4e86\u4ee4\u4eba\u96be\u4ee5\u7f6e\u4fe1\u7684\u6027\u80fd\u6c34\u5e73\uff0c\u7f29\u5c0f\u4e86\u95ed\u6e90\u548c\u5f00\u6e90\u6a21\u578b\u4e4b\u95f4\u7684\u5dee\u8ddd\u3002\u60a8\u53ef\u4ee5\u9488\u5bf9\u7279\u5b9a\u7528\u4f8b\u5fae\u8c03 Llama 3.1\uff0c\u800c\u4e0d\u662f\u4f7f\u7528\u51bb\u7ed3\u7684\u901a\u7528 LLM\uff0c\u4f8b\u5982 GPT-4o \u548c Claude 3.5\uff0c\u4ee5\u66f4\u4f4e\u7684\u6210\u672c\u5b9e\u73b0\u66f4\u597d\u7684\u6027\u80fd\u548c\u53ef\u5b9a\u5236\u6027\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u5c06\u5bf9\u76d1\u7763\u5fae\u8c03\u8fdb\u884c\u5168\u9762\u6982\u8ff0\u3002\u6211\u4eec\u5c06\u5c06\u5176\u4e0e\u63d0\u793a\u5de5\u7a0b\u8fdb\u884c\u6bd4\u8f83\uff0c\u4ee5\u4e86\u89e3\u4f55\u65f6\u4f7f\u7528\u5b83\u6709\u610f\u4e49\uff0c \u8be6\u7ec6\u8bf4\u660e\u4e3b\u8981\u6280\u672f\u53ca\u5176\u4f18\u7f3a\u70b9\uff0c \u5e76\u4ecb\u7ecd\u4e3b\u8981\u6982\u5ff5\uff0c \u4f8b\u5982 LoRA \u8d85\u53c2\u6570\uff0c \u5b58\u50a8\u683c\u5f0f\uff0c \u548c\u804a\u5929\u6a21\u677f.\u6700\u540e\uff0c\u6211\u4eec\u5c06\u901a\u8fc7\u5728 Google Colab \u4e2d\u4f7f\u7528 Unsloth \u8fdb\u884c\u6700\u5148\u8fdb\u7684\u4f18\u5316\u6765\u5fae\u8c03 Llama 3.1 8B \u6765\u5728\u5b9e\u8df5\u4e2d\u5b9e\u73b0\u5b83\u3002<\/p>\n\n\n\n

\u672c\u6587\u4e2d\u4f7f\u7528\u7684\u6240\u6709\u4ee3\u7801\u90fd\u53ef\u4ee5\u5728 Google Colab<\/a> \u548c LLM \u8bfe\u7a0b<\/a>\u4e2d\u627e\u5230\u3002\u7279\u522b\u611f\u8c22 Daniel Han \u56de\u7b54\u6211\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n

<\/a>\ud83d\udd27 \u76d1\u7763\u5fae\u8c03<\/h2>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u76d1\u7763\u5fae\u8c03 \uff08SFT\uff09 \u662f\u4e00\u79cd\u6539\u8fdb\u548c\u5b9a\u5236<\/strong>\u9884\u8bad\u7ec3 LLM \u7684\u65b9\u6cd5\u3002\u5b83\u6d89\u53ca\u5728\u8f83\u5c0f\u7684\u6307\u4ee4\u548c\u7b54\u6848\u6570\u636e\u96c6\u4e0a\u91cd\u65b0\u8bad\u7ec3\u57fa\u7840\u6a21\u578b\u3002\u4e3b\u8981\u76ee\u6807\u662f\u5c06\u9884\u6d4b\u6587\u672c\u7684\u57fa\u672c\u6a21\u578b\u8f6c\u6362\u4e3a\u53ef\u4ee5\u9075\u5faa\u6307\u793a\u5e76\u56de\u7b54\u95ee\u9898\u7684\u52a9\u624b\u3002SFT \u8fd8\u53ef\u4ee5\u589e\u5f3a\u6a21\u578b\u7684\u6574\u4f53\u6027\u80fd\u3001\u6dfb\u52a0\u65b0\u77e5\u8bc6\u6216\u4f7f\u5176\u9002\u5e94\u7279\u5b9a\u4efb\u52a1\u548c\u9886\u57df\u3002\u7136\u540e\uff0c\u5fae\u8c03\u7684\u6a21\u578b\u53ef\u4ee5\u7ecf\u5386\u4e00\u4e2a\u53ef\u9009\u7684\u504f\u597d\u5bf9\u9f50\u9636\u6bb5\uff08\u8bf7\u53c2\u9605\u6211\u5173\u4e8e DPO \u7684\u6587\u7ae0<\/a>\uff09\uff0c\u4ee5\u5220\u9664\u4e0d\u9700\u8981\u7684\u54cd\u5e94\u3001\u4fee\u6539\u5176\u6837\u5f0f\u7b49\u3002<\/p>\n\n\n\n

\u4e0b\u56fe\u663e\u793a\u4e86\u4e00\u4e2a\u6307\u4ee4\u793a\u4f8b\u3002\u5b83\u5305\u62ec\u7528\u4e8e\u5f15\u5bfc\u6a21\u578b\u7684\u7cfb\u7edf\u63d0\u793a\u3001\u7528\u4e8e\u63d0\u4f9b\u4efb\u52a1\u7684\u7528\u6237\u63d0\u793a\u4ee5\u53ca\u6a21\u578b\u9884\u671f\u751f\u6210\u7684\u8f93\u51fa\u3002\u60a8\u53ef\u4ee5\u5728 LLM \u6570\u636e\u96c6 GitHub \u5b58\u50a8\u5e93\u4e2d\u627e\u5230\ud83d\udcbe<\/a>\u9ad8\u8d28\u91cf\u7684\u5f00\u6e90\u6307\u4ee4\u6570\u636e\u96c6\u5217\u8868\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u5728\u8003\u8651 SFT \u4e4b\u524d\uff0c\u6211\u5efa\u8bae\u5c1d\u8bd5\u63d0\u793a\u5de5\u7a0b\u6280\u672f\uff0c\u4f8b\u5982\u5c11\u6837\u672c\u63d0\u793a<\/strong>\u6216\u68c0\u7d22\u589e\u5f3a\u751f\u6210<\/strong> \uff08RAG\uff09\u3002\u5728\u5b9e\u8df5\u4e2d\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u53ef\u4ee5\u89e3\u51b3\u8bb8\u591a\u95ee\u9898\uff0c\u800c\u65e0\u9700\u5fae\u8c03\uff0c\u4f7f\u7528\u95ed\u6e90\u6216\u5f00\u653e\u6743\u91cd\u6a21\u578b\uff08\u4f8b\u5982\uff0cLlama 3.1 Instruct\uff09\u3002\u5982\u679c\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u80fd\u6ee1\u8db3\u60a8\u7684\u76ee\u6807\uff08\u5728\u8d28\u91cf\u3001\u6210\u672c\u3001\u5ef6\u8fdf\u7b49\u65b9\u9762\uff09\uff0c\u90a3\u4e48\u5f53\u6307\u4ee4\u6570\u636e\u53ef\u7528\u65f6\uff0cSFT \u5c31\u6210\u4e3a\u4e00\u4e2a\u53ef\u884c\u7684\u9009\u62e9\u3002\u8bf7\u6ce8\u610f\uff0cSFT \u8fd8\u63d0\u4f9b\u989d\u5916\u7684\u63a7\u5236\u548c\u53ef\u5b9a\u5236\u6027\u7b49\u4f18\u52bf\uff0c\u4ee5\u521b\u5efa\u4e2a\u6027\u5316\u7684 LLM\u3002<\/p>\n\n\n\n

\u4f46\u662f\uff0cSFT \u6709\u5c40\u9650\u6027\u3002\u5728\u5229\u7528\u57fa\u7840\u6a21\u578b\u4e2d\u5df2\u6709\u7684\u77e5\u8bc6\u65f6\uff0c\u5b83\u6548\u679c\u6700\u597d\u3002\u5b66\u4e60\u5168\u65b0\u7684\u4fe1\u606f\uff08\u5982\u672a\u77e5\u8bed\u8a00\uff09\u53ef\u80fd\u5177\u6709\u6311\u6218\u6027\uff0c\u5e76\u5bfc\u81f4\u66f4\u9891\u7e41\u7684\u5e7b\u89c9\u3002\u5bf9\u4e8e\u57fa\u7840\u6a21\u578b\u672a\u77e5\u7684\u65b0\u57df\uff0c\u5efa\u8bae\u5148\u5728\u539f\u59cb\u6570\u636e\u96c6\u4e0a\u6301\u7eed\u9884\u8bad\u7ec3\u5b83\u3002<\/p>\n\n\n\n

\u53e6\u4e00\u65b9\u9762\uff0c\u6307\u4ee4\u6a21\u578b\uff08\u5373\u5df2\u7ecf\u5fae\u8c03\u7684\u6a21\u578b\uff09\u5df2\u7ecf\u975e\u5e38\u63a5\u8fd1\u60a8\u7684\u9700\u6c42\u3002\u4f8b\u5982\uff0c\u4e00\u4e2a\u6a21\u578b\u53ef\u80fd\u8868\u73b0\u975e\u5e38\u597d\uff0c\u4f46\u58f0\u660e\u5b83\u662f\u7531 OpenAI \u6216 Meta \u800c\u4e0d\u662f\u60a8\u8bad\u7ec3\u7684\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u60a8\u53ef\u80fd\u5e0c\u671b\u4f7f\u7528\u504f\u597d\u5bf9\u9f50\u6765\u7a0d\u5fae\u5f15\u5bfc\u6307\u793a\u6a21\u578b\u7684\u884c\u4e3a\u3002\u901a\u8fc7\u4e3a\u4e00\u5c0f\u7ec4\u6307\u4ee4\uff08100 \u5230 1000 \u4e2a\u6837\u672c\uff09\u63d0\u4f9b\u9009\u62e9\u548c\u62d2\u7edd\u7684\u6837\u672c\uff0c\u60a8\u53ef\u4ee5\u5f3a\u5236 LLM \u8bf4\u60a8\u8bad\u7ec3\u4e86\u5b83\u800c\u4e0d\u662f OpenAI\u3002<\/p>\n\n\n\n

<\/a>\u2696\ufe0f SFT\u6280\u672f<\/h2>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u4e09\u79cd\u6700\u6d41\u884c\u7684 SFT \u6280\u672f\u662f\u5b8c\u5168\u5fae\u8c03\u3001LoRA \u548c QLoRA\u3002<\/p>\n\n\n\n

\u5b8c\u5168\u5fae\u8c03<\/strong>\u662f\u6700\u76f4\u63a5\u7684 SFT \u6280\u672f\u3002\u5b83\u6d89\u53ca\u5728\u6307\u4ee4\u6570\u636e\u96c6\u4e0a\u91cd\u65b0\u8bad\u7ec3\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u6240\u6709\u53c2\u6570\u3002\u8fd9\u79cd\u65b9\u6cd5\u901a\u5e38\u63d0\u4f9b\u6700\u4f73\u7ed3\u679c\uff0c\u4f46\u9700\u8981\u5927\u91cf\u7684\u8ba1\u7b97\u8d44\u6e90\uff08\u9700\u8981\u51e0\u4e2a\u9ad8\u7aef GPU \u6765\u5fae\u8c03 8B \u6a21\u578b\uff09\u3002\u56e0\u4e3a\u5b83\u4fee\u6539\u4e86\u6574\u4e2a\u6a21\u578b\uff0c\u6240\u4ee5\u5b83\u4e5f\u662f\u6700\u5177\u7834\u574f\u6027\u7684\u65b9\u6cd5\uff0c\u5e76\u53ef\u80fd\u5bfc\u81f4\u707e\u96be\u6027\u5730\u5fd8\u8bb0\u4ee5\u524d\u7684\u6280\u80fd\u548c\u77e5\u8bc6\u3002<\/p>\n\n\n\n

\u4f4e\u79e9\u81ea\u9002\u5e94 \uff08LoRA\uff09<\/strong> \u662f\u4e00\u79cd\u6d41\u884c\u7684\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u6280\u672f\u3002\u5b83\u4e0d\u662f\u91cd\u65b0\u8bad\u7ec3\u6574\u4e2a\u6a21\u578b\uff0c\u800c\u662f\u51bb\u7ed3\u6743\u91cd\u5e76\u5728\u6bcf\u4e2a\u76ee\u6807\u5c42\u5f15\u5165\u5c0f\u9002\u914d\u5668\uff08\u4f4e\u79e9\u77e9\u9635\uff09\u3002\u8fd9\u4f7f\u5f97 LoRA \u80fd\u591f\u8bad\u7ec3\u8bb8\u591a\u53c2\u6570\uff0c\u8fd9\u4e9b\u53c2\u6570\u5927\u5927\u4f4e\u4e8e\u5b8c\u5168\u5fae\u8c03 \uff08\u5c0f\u4e8e 1%\uff09\uff0c \u51cf\u5c11\u4e86\u5185\u5b58\u4f7f\u7528\u548c\u8bad\u7ec3\u65f6\u95f4.\u8fd9\u79cd\u65b9\u6cd5\u662f\u975e\u7834\u574f\u6027\u7684\uff0c\u56e0\u4e3a\u539f\u59cb\u53c2\u6570\u88ab\u51bb\u7ed3\uff0c\u7136\u540e\u53ef\u4ee5\u968f\u610f\u5207\u6362\u6216\u7ec4\u5408\u9002\u914d\u5668\u3002<\/p>\n\n\n\n

QLoRA\uff08\u91cf\u5316\u611f\u77e5\u4f4e\u79e9\u9002\u5e94\uff09<\/strong>\u662f LoRA \u7684\u6269\u5c55\uff0c\u53ef\u8282\u7701\u66f4\u5927\u7684\u5185\u5b58\u3002\u4e0e\u6807\u51c6 LoRA \u76f8\u6bd4\uff0c\u5b83\u53ef\u989d\u5916\u51cf\u5c11\u9ad8\u8fbe 33% \u7684\u5185\u5b58\uff0c\u56e0\u6b64\u5728 GPU \u5185\u5b58\u53d7\u9650\u65f6\u7279\u522b\u6709\u7528\u3002\u8fd9\u79cd\u6548\u7387\u7684\u63d0\u9ad8\u662f\u4ee5\u66f4\u957f\u7684\u8bad\u7ec3\u65f6\u95f4\u4e3a\u4ee3\u4ef7\u7684\uff0cQLoRA \u7684\u8bad\u7ec3\u65f6\u95f4\u901a\u5e38\u6bd4\u5e38\u89c4 LoRA \u591a 39%\u3002<\/p>\n\n\n\n

\u867d\u7136 QLoRA \u9700\u8981\u66f4\u591a\u7684\u8bad\u7ec3\u65f6\u95f4\uff0c\u4f46\u5176\u5927\u91cf\u5185\u5b58\u8282\u7701\u53ef\u4ee5\u4f7f\u5176\u6210\u4e3a GPU \u5185\u5b58\u6709\u9650\u7684\u573a\u666f\u4e2d\u552f\u4e00\u53ef\u884c\u7684\u9009\u62e9\u3002\u51fa\u4e8e\u8fd9\u4e2a\u539f\u56e0\uff0c\u6211\u4eec\u5c06\u5728\u4e0b\u4e00\u8282\u4e2d\u4f7f\u7528\u8fd9\u79cd\u6280\u672f\u6765\u5fae\u8c03 Google Colab \u4e0a\u7684 Llama 3.1 8B \u6a21\u578b\u3002<\/p>\n\n\n\n

<\/a>\ud83e\udd99 \u5fae\u8c03 Llama 3.1 8B<\/h2>\n\n\n\n

\u4e3a\u4e86\u6709\u6548\u5730\u5fae\u8c03 Llama 3.1 8B<\/a> \u6a21\u578b\uff0c\u6211\u4eec\u5c06\u4f7f\u7528 Daniel \u548c Michael Han \u7684 Unsloth<\/a> \u5e93\u3002\u5f97\u76ca\u4e8e\u5176\u81ea\u5b9a\u4e49\u5185\u6838\uff0c\u4e0e\u5176\u4ed6\u9009\u9879\u76f8\u6bd4\uff0cUnsloth \u7684\u8bad\u7ec3\u901f\u5ea6\u63d0\u9ad8\u4e86 2 \u500d\uff0c\u5185\u5b58\u4f7f\u7528\u7387\u63d0\u9ad8\u4e86 60%\uff0c\u975e\u5e38\u9002\u5408 Colab \u7b49\u53d7\u9650\u73af\u5883\u3002\u4e0d\u5e78\u7684\u662f\uff0cUnsloth \u76ee\u524d\u4ec5\u652f\u6301\u5355 GPU \u8bbe\u7f6e\u3002\u5bf9\u4e8e\u591a GPU \u8bbe\u7f6e\uff0c\u6211\u63a8\u8350\u6d41\u884c\u7684\u66ff\u4ee3\u54c1\uff0c\u5982 TRL<\/a> \u548c Axolotl<\/a>\uff08\u4e24\u8005\u90fd\u5305\u62ec Unsloth \u4f5c\u4e3a\u540e\u7aef\uff09\u3002<\/p>\n\n\n\n

\u5728\u6b64\u793a\u4f8b\u4e2d\uff0c\u6211\u4eec\u5c06\u5728 mlabonne\/FineTome-100k<\/a> \u6570\u636e\u96c6\u4e0a\u5bf9\u5176\u8fdb\u884c QLoRA \u5fae\u8c03\u3002\u8fd9\u662f\u6211\u4f7f\u7528 HuggingFaceFW\/fineweb-edu-classifier<\/a> \u91cd\u65b0\u8fc7\u6ee4\u7684 arcee-ai\/The-Tome<\/a>\uff08\u6ca1\u6709 arcee-ai\/qwen2-72b-magpie-en<\/a>\uff09\u7684\u5b50\u96c6\u3002\u8bf7\u6ce8\u610f\uff0c\u6b64\u5206\u7c7b\u5668\u4e0d\u662f\u4e3a\u6307\u4ee4\u6570\u636e\u8d28\u91cf\u8bc4\u4f30\u800c\u8bbe\u8ba1\u7684\uff0c\u4f46\u6211\u4eec\u53ef\u4ee5\u5c06\u5176\u7528\u4f5c\u7c97\u7565\u7684\u4ee3\u7406\u3002\u7531\u6b64\u4ea7\u751f\u7684 FineTome \u662f\u4e00\u4e2a\u8d85\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u96c6\uff0c\u5305\u62ec\u5bf9\u8bdd\u3001\u63a8\u7406\u95ee\u9898\u3001\u51fd\u6570\u8c03\u7528\u7b49\u3002<\/p>\n\n\n\n

\u8ba9\u6211\u4eec\u5148\u521b\u5efa\u73af\u5883\uff0c\u6d4b\u8bd5\u73af\u5883\u4e3a WSL2,Ubuntu 22.04.3 LTS\uff1a<\/p>\n\n\n\n

conda create -yn unsloth python=3.10\nconda activate unsloth\n<\/pre><\/div>\n\n\n\n
\u5b89\u88c5\u6240\u6709\u5fc5\u9700\u7684\u5e93\uff1a<\/p>\n\n\n\n
pip install \"unsloth[colab-new] @ git+https:\/\/github.com\/unslothai\/unsloth.git\"\npip install --no-deps \"xformers<0.0.27\" \"trl<0.9.0\" peft accelerate bitsandbytes\npip install torch==2.3.0+cu121 torchvision torchaudio --index-url https:\/\/download.pytorch.org\/whl\/cu121\npip install numpy==1.24.4\n<\/pre><\/div>\n\n\n\n
\u5b89\u88c5\u540e\uff0c\u6211\u4eec\u53ef\u4ee5\u6309\u4ee5\u4e0b\u6b65\u9aa4\u5bfc\u5165\u5b83\u4eec\u3002<\/p>\n\n\n\n
import torch\nfrom trl import SFTTrainer\nfrom datasets import load_dataset\nfrom transformers import TrainingArguments, TextStreamer\nfrom unsloth.chat_templates import get_chat_template\nfrom unsloth import FastLanguageModel, is_bfloat16_supported\n<\/pre><\/div>\n\n\n\n
\u73b0\u5728\u8ba9\u6211\u4eec\u52a0\u8f7d\u6a21\u578b\u3002\u7531\u4e8e\u6211\u4eec\u60f3\u4f7f\u7528 QLoRA\uff0c\u6211\u9009\u62e9\u4e86\u9884\u91cf\u5316\u7684 unsloth\/Meta-Llama-3.1-8B-bnb-4bit<\/a>\u3002\u4e0e\u539f\u59cb\u7684 16 \u4f4d\u7cbe\u5ea6\u6a21\u578b \uff0816 GB\uff09 \u76f8\u6bd4\uff0c\u8fd9\u6b3e meta-llama\/Meta-Llama-3.1-8B<\/a> \u7684 4 \u4f4d\u7cbe\u5ea6\u7248\u672c\u660e\u663e\u66f4\u5c0f \uff085.4 GB\uff09\uff0c\u4e0b\u8f7d\u901f\u5ea6\u66f4\u5feb\u3002\u6211\u4eec\u4f7f\u7528 bitsandbytes \u5e93\u4ee5 NF4 \u683c\u5f0f\u52a0\u8f7d\u3002<\/p>\n\n\n\n
\u52a0\u8f7d\u6a21\u578b\u65f6\uff0c\u6211\u4eec\u5fc5\u987b\u6307\u5b9a\u6700\u5927\u5e8f\u5217\u957f\u5ea6\uff0c\u8fd9\u4f1a\u9650\u5236\u5176\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002Llama 3.1 \u652f\u6301\u9ad8\u8fbe 128k \u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u4f46\u5728\u6b64\u793a\u4f8b\u4e2d\uff0c\u6211\u4eec\u5c06\u5c06\u5176\u8bbe\u7f6e\u4e3a 2,048\uff0c\u56e0\u4e3a\u5b83\u6d88\u8017\u66f4\u591a\u7684\u8ba1\u7b97\u548c VRAM\u3002\u6700\u540e\uff0cdtype<\/code> \u53c2\u6570\u4f1a\u81ea\u52a8\u68c0\u6d4b\u60a8\u7684 GPU \u662f\u5426\u652f\u6301 BF16 \u683c\u5f0f<\/a>\uff0c\u4ee5\u4fbf\u5728\u8bad\u7ec3\u671f\u95f4\u63d0\u9ad8\u7a33\u5b9a\u6027\uff08\u6b64\u529f\u80fd\u4ec5\u9650\u4e8e Ampere \u548c\u66f4\u65b0\u7684 GPU\uff09\u3002<\/p>\n\n\n\n
max_seq_length = 2048\nmodel, tokenizer = FastLanguageModel.from_pretrained(\n    model_name=\"unsloth\/Meta-Llama-3.1-8B-bnb-4bit\",\n    max_seq_length=max_seq_length,\n    load_in_4bit=True,\n    dtype=None,\n)\n<\/pre><\/div>\n\n\n\n
\u73b0\u5728\u6211\u4eec\u7684\u6a21\u578b\u5df2\u4ee5 4 \u4f4d\u7cbe\u5ea6\u52a0\u8f7d\uff0c\u6211\u4eec\u5e0c\u671b\u4e3a\u4f7f\u7528 LoRA \u9002\u914d\u5668\u8fdb\u884c\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u505a\u597d\u51c6\u5907\u3002LoRA \u6709\u4e09\u4e2a\u91cd\u8981\u53c2\u6570\uff1a<\/p>\n\n\n\n