[{"data":1,"prerenderedAt":1135},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"wiki-lora":9},{"tools":4,"reviews":5},65,7,{"tools":4,"reviews":5,"playbooks":7,"news":8},10,8,{"id":10,"title":11,"body":12,"category":1118,"description":72,"extension":1119,"meta":1120,"navigation":729,"path":1121,"published":1122,"relatedModels":1123,"relatedTools":1126,"seo":1127,"slug":1128,"stem":1129,"summary":1130,"tags":1131,"updated":1122,"__hash__":1134},"wiki\u002Fwiki\u002Flora.md","LoRA（低秩适配）",{"type":13,"value":14,"toc":1088},"minimark",[15,20,29,32,35,39,42,55,59,62,73,76,84,89,92,96,151,155,158,162,165,176,180,183,187,190,289,295,298,302,305,309,312,344,348,351,403,407,410,482,489,493,496,553,556,560,563,644,650,654,657,663,670,681,684,687,691,694,819,823,965,969,972,976,1051,1054,1084],[16,17,19],"h2",{"id":18},"什么是-lora","什么是 LoRA",[21,22,23,24,28],"p",{},"LoRA（Low-Rank Adaptation）是一种参数高效微调方法。核心思想：",[25,26,27],"strong",{},"不修改原始模型权重，而是在旁边加一个很小的\"适配器\"矩阵来学习任务特定知识","。",[21,30,31],{},"打个比方：原模型是一本教科书（不能改写），LoRA 是你在书页边写的笔记——不改原书，但补充了特定场景的知识。",[16,33,34],{"id":34},"原理",[36,37,38],"h3",{"id":38},"传统微调",[21,40,41],{},"全量微调要更新模型所有参数。一个 70B 模型有 700 亿参数，全部更新需要：",[43,44,45,49,52],"ul",{},[46,47,48],"li",{},"700 亿参数的梯度计算",[46,50,51],{},"700 亿参数的优化器状态",[46,53,54],{},"至少 8×A100 GPU",[36,56,58],{"id":57},"lora-微调","LoRA 微调",[21,60,61],{},"LoRA 的数学原理：大模型权重矩阵 W 的更新可以用两个小矩阵 A × B 来近似：",[63,64,69],"pre",{"className":65,"code":67,"language":68},[66],"language-text","W' = W + A × B\n\nW: 原始权重 (d×d)，冻结不动\nA: 降维矩阵 (d×r)，需要训练\nB: 升维矩阵 (r×d)，需要训练\nr: 秩（rank），通常 8-64\n","text",[70,71,67],"code",{"__ignoreMap":72},"",[21,74,75],{},"当 r=8, d=4096 时：",[43,77,78,81],{},[46,79,80],{},"原始参数：4096×4096 = 1678 万",[46,82,83],{},"LoRA 参数：4096×8 + 8×4096 = 6.6 万",[21,85,86,28],{},[25,87,88],{},"训练参数减少 99.6%",[16,90,91],{"id":91},"优势",[36,93,95],{"id":94},"_1-显存大幅降低","1. 显存大幅降低",[97,98,99,115],"table",{},[100,101,102],"thead",{},[103,104,105,109,112],"tr",{},[106,107,108],"th",{},"方法",[106,110,111],{},"70B 模型显存",[106,113,114],{},"GPU 需求",[116,117,118,130,140],"tbody",{},[103,119,120,124,127],{},[121,122,123],"td",{},"全量微调",[121,125,126],{},"~500GB",[121,128,129],{},"8×A100 80G",[103,131,132,134,137],{},[121,133,58],{},[121,135,136],{},"~80GB",[121,138,139],{},"1×A100 80G",[103,141,142,145,148],{},[121,143,144],{},"QLoRA（量化+LoRA）",[121,146,147],{},"~24GB",[121,149,150],{},"1×RTX 4090",[36,152,154],{"id":153},"_2-训练速度快","2. 训练速度快",[21,156,157],{},"参数少，梯度计算和优化器更新都更快。通常比全量微调快 2-3 倍。",[36,159,161],{"id":160},"_3-可插拔","3. 可插拔",[21,163,164],{},"LoRA 适配器是一个独立的小文件（几十 MB），可以：",[43,166,167,170,173],{},[46,168,169],{},"随时加载\u002F卸载",[46,171,172],{},"多个 LoRA 切换使用",[46,174,175],{},"不同用户用不同 LoRA",[36,177,179],{"id":178},"_4-不破坏原模型","4. 不破坏原模型",[21,181,182],{},"原模型权重不变，可以随时回到原始状态。多个 LoRA 可以叠加。",[16,184,186],{"id":185},"lora-家族lora-qlora-dora-adalora","LoRA 家族：LoRA \u002F QLoRA \u002F DoRA \u002F AdaLoRA",[21,188,189],{},"LoRA 出来后衍生了一堆变体，按需选用：",[97,191,192,207],{},[100,193,194],{},[103,195,196,198,201,204],{},[106,197,108],{},[106,199,200],{},"核心改动",[106,202,203],{},"相对 LoRA 的变化",[106,205,206],{},"适合",[116,208,209,225,241,257,273],{},[103,210,211,216,219,222],{},[121,212,213],{},[25,214,215],{},"LoRA",[121,217,218],{},"基础版",[121,220,221],{},"—",[121,223,224],{},"通用、入门首选",[103,226,227,232,235,238],{},[121,228,229],{},[25,230,231],{},"QLoRA",[121,233,234],{},"把原模型量化到 4bit 再加 LoRA",[121,236,237],{},"显存 -70%，速度略慢",[121,239,240],{},"消费级 GPU",[103,242,243,248,251,254],{},[121,244,245],{},[25,246,247],{},"DoRA",[121,249,250],{},"分解为「方向 + 大小」分别学",[121,252,253],{},"同参数量下效果接近全量微调",[121,255,256],{},"追求效果上限",[103,258,259,264,267,270],{},[121,260,261],{},[25,262,263],{},"AdaLoRA",[121,265,266],{},"训练中动态调整每层的 rank",[121,268,269],{},"自动找最优 rank 分配",[121,271,272],{},"不想手调 rank",[103,274,275,280,283,286],{},[121,276,277],{},[25,278,279],{},"VeRA",[121,281,282],{},"共享随机矩阵 + 缩放向量",[121,284,285],{},"参数量再降 10x",[121,287,288],{},"多 LoRA 大量部署",[21,290,291,294],{},[25,292,293],{},"经验","：80% 场景 LoRA 就够，显存不够上 QLoRA。DoRA 适合 LoRA 效果不到位但又不想全量微调的中间地带。",[16,296,297],{"id":297},"劣势",[36,299,301],{"id":300},"_1-效果略逊全量微调","1. 效果略逊全量微调",[21,303,304],{},"LoRA 在大多数任务上接近全量微调，但在需要大幅改变模型行为的场景（如全新语言学习）可能不如全量。",[36,306,308],{"id":307},"_2-需要调参","2. 需要调参",[21,310,311],{},"LoRA 的关键参数：",[43,313,314,320,326,332,338],{},[46,315,316,319],{},[25,317,318],{},"r（秩）","：8-64，越大效果越好但参数越多",[46,321,322,325],{},[25,323,324],{},"alpha","：缩放因子，通常设为 r 的 2 倍",[46,327,328,331],{},[25,329,330],{},"target_modules","：对哪些层加 LoRA（通常选 q_proj, v_proj）",[46,333,334,337],{},[25,335,336],{},"dropout","：0.0-0.1，防过拟合",[46,339,340,343],{},[25,341,342],{},"学习率","：1e-4 到 5e-4（比全量微调高一个数量级）",[36,345,347],{"id":346},"_3-推理时需要决定合不合并","3. 推理时需要决定合不合并",[21,349,350],{},"LoRA 推理时有两种姿势，各有取舍：",[97,352,353,368],{},[100,354,355],{},[103,356,357,360,363,366],{},[106,358,359],{},"方案",[106,361,362],{},"推理性能",[106,364,365],{},"灵活性",[106,367,206],{},[116,369,370,387],{},[103,371,372,378,381,384],{},[121,373,374,377],{},[25,375,376],{},"合并到原模型","（merge）",[121,379,380],{},"与原模型完全一致，零开销",[121,382,383],{},"死板，换 LoRA 要重新合并、重新部署",[121,385,386],{},"单一固定 LoRA、生产稳定环境",[103,388,389,394,397,400],{},[121,390,391],{},[25,392,393],{},"运行时加载",[121,395,396],{},"多一次矩阵加法，延迟 +5-15%",[121,398,399],{},"同进程切换多个 LoRA、A\u002FB 测试",[121,401,402],{},"多租户 \u002F 多版本场景",[16,404,406],{"id":405},"target_modules-怎么选","target_modules 怎么选",[21,408,409],{},"这是 LoRA 调参里最玄学也最关键的一项。常见组合：",[97,411,412,428],{},[100,413,414],{},[103,415,416,419,422,425],{},[106,417,418],{},"组合",[106,420,421],{},"训练参数",[106,423,424],{},"效果",[106,426,427],{},"何时用",[116,429,430,447,464],{},[103,431,432,438,441,444],{},[121,433,434,435],{},"只 ",[70,436,437],{},"q_proj, v_proj",[121,439,440],{},"最少",[121,442,443],{},"一般",[121,445,446],{},"显存极紧 \u002F 快速实验",[103,448,449,455,458,461],{},[121,450,451,454],{},[70,452,453],{},"q_proj, k_proj, v_proj, o_proj","（全注意力）",[121,456,457],{},"中",[121,459,460],{},"较好",[121,462,463],{},"通用推荐起点",[103,465,466,473,476,479],{},[121,467,468,469,472],{},"全注意力 + MLP（",[70,470,471],{},"gate_proj, up_proj, down_proj","）",[121,474,475],{},"多",[121,477,478],{},"接近全量微调",[121,480,481],{},"任务复杂、显存够",[21,483,484,485,488],{},"LLaMA-Factory \u002F PEFT 都允许 ",[70,486,487],{},"target_modules=\"all-linear\""," 一把梭，对小模型（\u003C 13B）通常直接用这个最省心。",[16,490,492],{"id":491},"qlora4bit-量化-lora","QLoRA：4bit 量化 + LoRA",[21,494,495],{},"QLoRA 是 LoRA 的进一步优化：把原模型量化到 4bit 再做 LoRA 微调。",[97,497,498,509],{},[100,499,500],{},[103,501,502,505,507],{},[106,503,504],{},"指标",[106,506,215],{},[106,508,231],{},[116,510,511,521,531,542],{},[103,512,513,515,518],{},[121,514,111],{},[121,516,517],{},"80GB",[121,519,520],{},"24GB",[103,522,523,525,528],{},[121,524,114],{},[121,526,527],{},"A100",[121,529,530],{},"RTX 4090",[103,532,533,536,539],{},[121,534,535],{},"效果损失",[121,537,538],{},"无",[121,540,541],{},"\u003C1%",[103,543,544,547,550],{},[121,545,546],{},"训练速度",[121,548,549],{},"基准",[121,551,552],{},"慢 20%",[21,554,555],{},"QLoRA 让在消费级 GPU（RTX 4090\u002F3090）上微调 70B 模型成为可能。",[16,557,559],{"id":558},"推理服务的-lora-支持","推理服务的 LoRA 支持",[21,561,562],{},"把 LoRA 跑到生产，关键看推理框架支不支持多 LoRA 动态切换：",[97,564,565,578],{},[100,566,567],{},[103,568,569,572,575],{},[106,570,571],{},"框架",[106,573,574],{},"LoRA 支持",[106,576,577],{},"多 LoRA 热加载",[116,579,580,593,606,618,631],{},[103,581,582,587,590],{},[121,583,584],{},[25,585,586],{},"vLLM",[121,588,589],{},"✅",[121,591,592],{},"✅ 一份基础模型 + 多个 LoRA 同时服务",[103,594,595,601,603],{},[121,596,597,600],{},[25,598,599],{},"TGI","（HuggingFace）",[121,602,589],{},[121,604,605],{},"✅ 支持运行时切换",[103,607,608,613,615],{},[121,609,610],{},[25,611,612],{},"TensorRT-LLM",[121,614,589],{},[121,616,617],{},"⚠️ 需要预编译",[103,619,620,625,628],{},[121,621,622],{},[25,623,624],{},"llama.cpp",[121,626,627],{},"✅ GGUF + LoRA",[121,629,630],{},"⚠️ 通常建议合并后再用",[103,632,633,638,641],{},[121,634,635],{},[25,636,637],{},"Ollama",[121,639,640],{},"⚠️ 通常合并",[121,642,643],{},"❌",[21,645,646,649],{},[25,647,648],{},"真正的杀手锏场景","：vLLM 的多 LoRA 部署。1 张 80G A100 可以同时服务 1 个 70B 基础模型 + 几十个 LoRA Adapter，不同用户请求路由到不同 LoRA。SaaS 化 fine-tune 服务（Together、Modal 等）背后基本都是这套。",[16,651,653],{"id":652},"多-adapter-切换的实战案例","多 Adapter 切换的实战案例",[21,655,656],{},"一个客服系统服务多个品牌客户：",[63,658,661],{"className":659,"code":660,"language":68},[66],"共享基础模型: Qwen-2.5-7B\n  ├─ adapter_brand_A.safetensors (40MB)   ← A 品牌话术 \u002F 知识\n  ├─ adapter_brand_B.safetensors (40MB)   ← B 品牌话术 \u002F 知识\n  └─ adapter_brand_C.safetensors (40MB)\n",[70,662,660],{"__ignoreMap":72},[21,664,665,666,669],{},"请求时按 ",[70,667,668],{},"brand"," 字段路由到对应 LoRA。优势：",[43,671,672,675,678],{},[46,673,674],{},"7B 基础模型只占一份 GPU 显存",[46,676,677],{},"新增品牌只要训练一个 40MB 的 LoRA，几小时上线",[46,679,680],{},"任一品牌出问题，回滚单个 LoRA 即可，不影响其他",[21,682,683],{},"如果用全量微调，要么每个品牌一份 14GB 模型权重（贵），要么混在一起训（容易互相干扰）。",[16,685,686],{"id":686},"实操工具",[36,688,690],{"id":689},"llama-factory","LLaMA-Factory",[21,692,693],{},"开箱即用的 LoRA 微调工具：",[63,695,699],{"className":696,"code":697,"language":698,"meta":72,"style":72},"language-bash shiki shiki-themes github-light github-dark","# 安装\npip install llama-factory\n\n# Web UI 启动\nllamafactory-cli webui\n\n# 命令行微调 Qwen\nllamafactory-cli train \\\n  --model_name_or_path Qwen\u002FQwen2.5-7B \\\n  --finetuning_type lora \\\n  --lora_rank 8 \\\n  --dataset your_dataset \\\n  --output_dir .\u002Foutput\n","bash",[70,700,701,710,724,731,737,746,751,756,767,778,788,799,810],{"__ignoreMap":72},[702,703,706],"span",{"class":704,"line":705},"line",1,[702,707,709],{"class":708},"sJ8bj","# 安装\n",[702,711,713,717,721],{"class":704,"line":712},2,[702,714,716],{"class":715},"sScJk","pip",[702,718,720],{"class":719},"sZZnC"," install",[702,722,723],{"class":719}," llama-factory\n",[702,725,727],{"class":704,"line":726},3,[702,728,730],{"emptyLinePlaceholder":729},true,"\n",[702,732,734],{"class":704,"line":733},4,[702,735,736],{"class":708},"# Web UI 启动\n",[702,738,740,743],{"class":704,"line":739},5,[702,741,742],{"class":715},"llamafactory-cli",[702,744,745],{"class":719}," webui\n",[702,747,749],{"class":704,"line":748},6,[702,750,730],{"emptyLinePlaceholder":729},[702,752,753],{"class":704,"line":5},[702,754,755],{"class":708},"# 命令行微调 Qwen\n",[702,757,758,760,763],{"class":704,"line":8},[702,759,742],{"class":715},[702,761,762],{"class":719}," train",[702,764,766],{"class":765},"sj4cs"," \\\n",[702,768,770,773,776],{"class":704,"line":769},9,[702,771,772],{"class":765},"  --model_name_or_path",[702,774,775],{"class":719}," Qwen\u002FQwen2.5-7B",[702,777,766],{"class":765},[702,779,780,783,786],{"class":704,"line":7},[702,781,782],{"class":765},"  --finetuning_type",[702,784,785],{"class":719}," lora",[702,787,766],{"class":765},[702,789,791,794,797],{"class":704,"line":790},11,[702,792,793],{"class":765},"  --lora_rank",[702,795,796],{"class":765}," 8",[702,798,766],{"class":765},[702,800,802,805,808],{"class":704,"line":801},12,[702,803,804],{"class":765},"  --dataset",[702,806,807],{"class":719}," your_dataset",[702,809,766],{"class":765},[702,811,813,816],{"class":704,"line":812},13,[702,814,815],{"class":765},"  --output_dir",[702,817,818],{"class":719}," .\u002Foutput\n",[36,820,822],{"id":821},"peft-hugging-face","PEFT (Hugging Face)",[63,824,828],{"className":825,"code":826,"language":827,"meta":72,"style":72},"language-python shiki shiki-themes github-light github-dark","from peft import LoraConfig, get_peft_model\nfrom transformers import AutoModelForCausalLM\n\nmodel = AutoModelForCausalLM.from_pretrained(\"Qwen\u002FQwen2.5-7B\")\nconfig = LoraConfig(\n    r=8, lora_alpha=16,\n    target_modules=[\"q_proj\", \"v_proj\"],\n    lora_dropout=0.05,\n)\nmodel = get_peft_model(model, config)\n# 训练...\n","python",[70,829,830,846,858,862,879,889,914,935,947,951,960],{"__ignoreMap":72},[702,831,832,836,840,843],{"class":704,"line":705},[702,833,835],{"class":834},"szBVR","from",[702,837,839],{"class":838},"sVt8B"," peft ",[702,841,842],{"class":834},"import",[702,844,845],{"class":838}," LoraConfig, get_peft_model\n",[702,847,848,850,853,855],{"class":704,"line":712},[702,849,835],{"class":834},[702,851,852],{"class":838}," transformers ",[702,854,842],{"class":834},[702,856,857],{"class":838}," AutoModelForCausalLM\n",[702,859,860],{"class":704,"line":726},[702,861,730],{"emptyLinePlaceholder":729},[702,863,864,867,870,873,876],{"class":704,"line":733},[702,865,866],{"class":838},"model ",[702,868,869],{"class":834},"=",[702,871,872],{"class":838}," AutoModelForCausalLM.from_pretrained(",[702,874,875],{"class":719},"\"Qwen\u002FQwen2.5-7B\"",[702,877,878],{"class":838},")\n",[702,880,881,884,886],{"class":704,"line":739},[702,882,883],{"class":838},"config ",[702,885,869],{"class":834},[702,887,888],{"class":838}," LoraConfig(\n",[702,890,891,895,897,900,903,906,908,911],{"class":704,"line":748},[702,892,894],{"class":893},"s4XuR","    r",[702,896,869],{"class":834},[702,898,899],{"class":765},"8",[702,901,902],{"class":838},", ",[702,904,905],{"class":893},"lora_alpha",[702,907,869],{"class":834},[702,909,910],{"class":765},"16",[702,912,913],{"class":838},",\n",[702,915,916,919,921,924,927,929,932],{"class":704,"line":5},[702,917,918],{"class":893},"    target_modules",[702,920,869],{"class":834},[702,922,923],{"class":838},"[",[702,925,926],{"class":719},"\"q_proj\"",[702,928,902],{"class":838},[702,930,931],{"class":719},"\"v_proj\"",[702,933,934],{"class":838},"],\n",[702,936,937,940,942,945],{"class":704,"line":8},[702,938,939],{"class":893},"    lora_dropout",[702,941,869],{"class":834},[702,943,944],{"class":765},"0.05",[702,946,913],{"class":838},[702,948,949],{"class":704,"line":769},[702,950,878],{"class":838},[702,952,953,955,957],{"class":704,"line":7},[702,954,866],{"class":838},[702,956,869],{"class":834},[702,958,959],{"class":838}," get_peft_model(model, config)\n",[702,961,962],{"class":704,"line":790},[702,963,964],{"class":708},"# 训练...\n",[36,966,968],{"id":967},"axolotl","Axolotl",[21,970,971],{},"YAML 驱动的训练框架，社区配方丰富，适合追新（DoRA \u002F 新优化器先到）。",[16,973,975],{"id":974},"什么时候用-lora","什么时候用 LoRA",[97,977,978,988],{},[100,979,980],{},[103,981,982,985],{},[106,983,984],{},"场景",[106,986,987],{},"推荐",[116,989,990,998,1006,1014,1021,1028,1035,1043],{},[103,991,992,995],{},[121,993,994],{},"风格\u002F格式调整",[121,996,997],{},"✅ LoRA",[103,999,1000,1003],{},[121,1001,1002],{},"领域知识注入",[121,1004,1005],{},"✅ LoRA + RAG",[103,1007,1008,1011],{},[121,1009,1010],{},"多租户定制",[121,1012,1013],{},"✅ LoRA（多 Adapter 部署）",[103,1015,1016,1019],{},[121,1017,1018],{},"快速实验",[121,1020,997],{},[103,1022,1023,1025],{},[121,1024,240],{},[121,1026,1027],{},"✅ QLoRA",[103,1029,1030,1033],{},[121,1031,1032],{},"全新语言学习",[121,1034,123],{},[103,1036,1037,1040],{},[121,1038,1039],{},"安全对齐",[121,1041,1042],{},"全量微调 \u002F DPO",[103,1044,1045,1048],{},[121,1046,1047],{},"闭源模型微调",[121,1049,1050],{},"用厂商 fine-tuning API",[16,1052,1053],{"id":1053},"延伸阅读",[43,1055,1056,1064,1076],{},[46,1057,1058,1059],{},"路线选型：",[1060,1061,1063],"a",{"href":1062},"\u002Fwiki\u002Ffine-tuning-vs-rag.html","Fine-tuning vs RAG",[46,1065,1066,1067,1071,1072],{},"数据准备前：",[1060,1068,1070],{"href":1069},"\u002Fwiki\u002Fembedding.html","Embedding"," \u002F ",[1060,1073,1075],{"href":1074},"\u002Fwiki\u002Frag.html","RAG",[46,1077,1078,1079,1083],{},"模型部署：",[1060,1080,1082],{"href":1081},"\u002Fwiki\u002Ftoken.html","Token"," 计费视角",[1085,1086,1087],"style",{},"html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .s4XuR, html code.shiki .s4XuR{--shiki-default:#E36209;--shiki-dark:#FFAB70}",{"title":72,"searchDepth":726,"depth":726,"links":1089},[1090,1091,1095,1101,1102,1107,1108,1109,1110,1111,1116,1117],{"id":18,"depth":712,"text":19},{"id":34,"depth":712,"text":34,"children":1092},[1093,1094],{"id":38,"depth":726,"text":38},{"id":57,"depth":726,"text":58},{"id":91,"depth":712,"text":91,"children":1096},[1097,1098,1099,1100],{"id":94,"depth":726,"text":95},{"id":153,"depth":726,"text":154},{"id":160,"depth":726,"text":161},{"id":178,"depth":726,"text":179},{"id":185,"depth":712,"text":186},{"id":297,"depth":712,"text":297,"children":1103},[1104,1105,1106],{"id":300,"depth":726,"text":301},{"id":307,"depth":726,"text":308},{"id":346,"depth":726,"text":347},{"id":405,"depth":712,"text":406},{"id":491,"depth":712,"text":492},{"id":558,"depth":712,"text":559},{"id":652,"depth":712,"text":653},{"id":686,"depth":712,"text":686,"children":1112},[1113,1114,1115],{"id":689,"depth":726,"text":690},{"id":821,"depth":726,"text":822},{"id":967,"depth":726,"text":968},{"id":974,"depth":712,"text":975},{"id":1053,"depth":712,"text":1053},"methodology","md",{},"\u002Fwiki\u002Flora","2026-06-21",[1124,1125],"llama-4","qwen-3",null,{"title":11,"description":72},"lora","wiki\u002Flora","Low-Rank Adaptation，一种高效微调方法：冻结原模型权重，只训练一个很小的低秩矩阵，大幅降低微调成本。",[215,1132,1133],"微调","高效训练","-wdobqimIFHvkroMMQaVjM1ZIEzmf08L_2FwF_Qkkrc",1782316490750]