[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-peft-vs-full-fine-tuning-zh":3,"article-related-peft-vs-full-fine-tuning-zh":33,"series-industry-d1218662-3c24-4bd5-8fdd-826164864369":83},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":25,"views":29,"created_at":30,"published_at":31,"topic_cluster_id":32},"d1218662-3c24-4bd5-8fdd-826164864369","peft-vs-full-fine-tuning-zh","PEFT vs 全量微調","\u003Cp data-speakable=\"summary\">PEFT 多數情況是大型語言模型微調的預設選項，全量微調則適合需要更深層模型改動的少數情境。\u003C\u002Fp>\u003Cp>在 \u003Ca href=\"https:\u002F\u002Fbhavishyapandit9.substack.com\u002Fp\u002Fllm-fine-tuning-at-scale-peft-vs\">PEFT\u003C\u002Fa> 與全量微調之間做選擇，重點不是誰比較新，而是誰更符合你的 \u003Ca href=\"\u002Ftag\u002Fgpu\">GPU\u003C\u002Fa> 預算、上線方式與準確率目標。這篇是寫給正在決定要不要用 LoRA、QLoRA，還是直接把整個模型權重都打開來訓練的團隊。\u003C\u002Fp>\u003Ch2>一張表看懂\u003C\u002Fh2>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>維度\u003C\u002Fth>\u003Cth>\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.09685\">PEFT\u003C\u002Fa>\u003C\u002Fth>\u003Cth>全量微調\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>13B 模型可訓練參數\u003C\u002Ftd>\u003Ctd>LoRA r=16 約 2,600 萬，約 0.2%\u003C\u002Ftd>\u003Ctd>130 億，100%\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>13B 指令微調峰值 VRAM\u003C\u002Ftd>\u003Ctd>LoRA 約 56 GB，QLoRA 約 18 GB\u003C\u002Ftd>\u003Ctd>約 240 GB\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>單次訓練成本\u003C\u002Ftd>\u003Ctd>LoRA 約 US$28，QLoRA 約 US$20\u003C\u002Ftd>\u003Ctd>約 US$190，8×A100 80GB\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>與全量微調相比的準確率差距\u003C\u002Ftd>\u003Ctd>LoRA 約 0.2 到 0.8 個百分點，QLoRA 約 1.1 個百分點\u003C\u002Ftd>\u003Ctd>基準\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>每個客戶版本的部署體積\u003C\u002Ftd>\u003Ctd>Adapter 約 50 到 200 MB\u003C\u002Ftd>\u003Ctd>完整模型檢查點約 26 GB\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>推論延遲額外開銷\u003C\u002Ftd>\u003Ctd>合併後 0%，動態掛載約 5% 到 15%\u003C\u002Ftd>\u003Ctd>0%\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>PEFT\u003C\u002Fh2>\u003Cp>PEFT 的核心優勢是把「改模型」\u003Ca href=\"\u002Fnews\u002Fbtc-etf-outflows-turn-79k-into-line-to-watch-zh\">變成\u003C\u002Fa>「改一小段可訓練參數」。以 LoRA 來說，它是在凍結原始權重的前提下，加上一個低秩更新矩陣，所以訓練時真正要學的東西很少。對 13B 模型而言，r=16 大約只有 2,600 萬個可訓練參數，這種規模讓單卡微調不再只是理論上可行，而是實務上可操作。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780603379788-d2wm.png\" alt=\"PEFT vs 全量微調\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>QLoRA 進一步把基礎模型壓到 4-bit 儲存，再在訓練時動態還原並更新 bf16 adapter，讓原本需要大顯存的工作，能在 24 GB 等級的卡上跑起來。這也是為\u003Ca href=\"\u002Fnews\u002Fdenver-hailstorm-weather-infrastructure-risk-zh\">什麼\u003C\u002Fa>很多團隊把 PEFT 當成預設值，因為它不只省錢，還能把迭代速度、版本管理與多客戶部署一起簡化。\u003C\u002Fp>\u003Ch2>全量微調\u003C\u002Fh2>\u003Cp>全量微調的價值，在於它不是只學「補丁」，而是讓模型整體重新對齊任務。當你要做初始指令調教、擴展到新語言，或是想把模型行為往某個方向大幅推進時，更新全部權重會給你最大的可塑性。這種自由度常常是 PEFT 做不到的，尤其在你追求最後 1 到 2 個百分點的基準分數時。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780603385377-nsdc.png\" alt=\"PEFT vs 全量微調\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>但代價也最直接。13B 模型若以 fp16 儲存，光權重就已經約 26 GB，還沒算 optimizer states、gradient 與 activations。實際訓練時，記憶體需求很快就會衝到多卡等級，成本、排程與除錯都會一起變重。它適合研究型團隊、基礎模型團隊，或是確定需要深度改寫模型行為的案子，不適合頻繁重訓或要維護很多客製版本的產品環境。\u003C\u002Fp>\u003Ch2>差異不只在準確率\u003C\u002Fh2>\u003Cp>很多人第一眼只看 \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa>，但真正拉開差距的是營運方式。PEFT 的 adapter 通常只有 50 到 200 MB，代表你可以為不同客戶、不同部門各放一份版本，必要時還能熱切換。這種彈性對客服、法務、金融、內部知識助理這類需求很重要，因為它們常常不是要一個最強模型，而是要很多個穩定可控的模型變體。\u003C\u002Fp>\u003Cp>全量微調則比較像一次把整台機器重新校準。它的部署檔案大，回滾成本高，重新訓練也慢，但如果你的目標是把模型本體改到很深，像是改語言覆蓋、改 \u003Ca href=\"\u002Ftag\u002Ftoken\">token\u003C\u002Fa> 分佈，或把某些能力真正內化進去，這種做法仍然有它的必要性。換句話說，PEFT 解的是「\u003Ca href=\"\u002Fnews\u002Fhow-to-hire-mlops-engineer-2026-zh\">怎麼\u003C\u002Fa>快又省地適配」，全量微調解的是「怎麼真的改變模型」。\u003C\u002Fp>\u003Ch2>怎麼選\u003C\u002Fh2>\u003Cp>如果你是產品團隊、應用團隊，或手上有多個客戶版本要維護，先選 PEFT。它最適合那些已經有不錯基礎模型，只想把它調成更懂領域語言、更符合公司流程的讀者。你會得到較低的 GPU 壓力、較快的實驗週期，以及比較容易上線的部署方式。\u003C\u002Fp>\u003Cp>如果你是研究團隊、模型團隊，或你明確知道自己要改的是模型核心行為，就選全量微調。它比較適合追求極限準確率、要做基礎能力變更，或需要把模型往新語言與新模態推進的讀者。你要能接受更高的算力成本與更複雜的訓練流程。\u003C\u002Fp>\u003Cp>如果你的首要目標是省錢、快迭代、好部署，PEFT 幾乎是預設答案。只有在你真的需要深度改寫模型，而不是單純套用任務適配時，全量微調才會翻盤。\u003C\u002Fp>\u003Cp>預設推薦 PEFT，尤其是 LoRA 或 QLoRA；唯一會讓答案改變的情境，是你正在做基礎模型級別的訓練，或需要明顯改動模型本體行為。\u003C\u002Fp>","PEFT 適合多數大型語言模型微調情境，全量微調則適合需要深度改動模型行為的少數案例。","bhavishyapandit9.substack.com","https:\u002F\u002Fbhavishyapandit9.substack.com\u002Fp\u002Fllm-fine-tuning-at-scale-peft-vs",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780603379788-d2wm.png","industry","zh","2a33bea3-0362-4c05-90c8-181ad6ff11b9",[17,18,19,20,21,22,23,24],"PEFT","全量微調","LoRA","QLoRA","大型語言模型","微調","GPU","模型部署",[26,27,28],"PEFT 省顯存、省成本，適合多數應用型微調與多版本部署。","全量微調能給模型更大改動空間，但訓練與部署成本都高很多。","若目標是快速適配與低風險上線，先選 PEFT；若要深度改寫模型，才考慮全量微調。",0,"2026-06-04T20:02:31.805871+00:00","2026-06-04T20:02:31.782+00:00","fe20f6f6-432b-47bf-a410-a5f516d885ed",{"tags":34,"relatedLang":43,"relatedPosts":47},[35,37,39,40,42],{"name":20,"slug":36},"qlora",{"name":19,"slug":38},"lora",{"name":21,"slug":21},{"name":17,"slug":41},"peft",{"name":18,"slug":18},{"id":15,"slug":44,"title":45,"language":46},"peft-vs-full-fine-tuning-en","PEFT vs Full Fine-Tuning","en",[48,54,60,65,71,77],{"id":49,"slug":50,"title":51,"cover_image":52,"image_url":52,"created_at":53,"category":13},"58fe51d5-e1c0-4b6d-9033-c40eb1f4f811","efrain-juarez-player-to-liga-mx-coach-zh","Efraín Juárez：從球員到Liga MX教練","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780606983453-d55c.png","2026-06-04T21:02:35.135418+00:00",{"id":55,"slug":56,"title":57,"cover_image":58,"image_url":58,"created_at":59,"category":13},"791faf8a-031f-4843-856a-2fe1dd7bef11","denver-hailstorm-weather-infrastructure-risk-zh","為什麼丹佛冰雹提醒我們：天氣就是基礎設施風險","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780601581653-sptw.png","2026-06-04T19:32:32.181958+00:00",{"id":61,"slug":62,"title":63,"cover_image":11,"image_url":11,"created_at":64,"category":13},"4923364e-f9c3-42fc-ae92-89ee5a822575","how-to-hire-mlops-engineer-2026-zh","怎麼招到 MLOps 工程師","2026-06-04T19:17:26.372485+00:00",{"id":66,"slug":67,"title":68,"cover_image":69,"image_url":69,"created_at":70,"category":13},"f8ff68f8-1cca-4db8-b871-c7b0fdf8eeb5","4-takeaways-from-cloudflares-ai-first-reset-zh","4 個關於 Cloudflare AI-first 重整的重點","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780589879946-i7e3.png","2026-06-04T16:17:28.780759+00:00",{"id":72,"slug":73,"title":74,"cover_image":75,"image_url":75,"created_at":76,"category":13},"f1822ffc-fbe1-4c5f-aa5d-e6dc37ae54a5","5-ways-harriet-sperling-echoes-kate-middleton-zh","5 種 Harriet Sperling 與凱特王妃的相似之處","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780587192008-dmzo.png","2026-06-04T15:32:45.790575+00:00",{"id":78,"slug":79,"title":80,"cover_image":81,"image_url":81,"created_at":82,"category":13},"247a7941-89df-44fe-87d8-7e906dba45f3","5-kops-release-notes-for-kubernetes-admins-zh","5 個 kOps 版本重點","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780586284483-09lo.png","2026-06-04T15:17:30.309022+00:00",[84,89,94,99,104,109,114,119,124,129],{"id":85,"slug":86,"title":87,"created_at":88},"ee073da7-28b3-4752-a319-5a501459fb87","ai-in-2026-what-actually-matters-now-zh","2026 AI 真正重要的事","2026-03-26T07:09:12.008134+00:00",{"id":90,"slug":91,"title":92,"created_at":93},"83bd1795-8548-44c9-9a7e-de50a0923f71","trump-ai-framework-power-speech-state-preemption-zh","川普 AI 框架瞄準電力、言論與州權","2026-03-26T07:12:18.695466+00:00",{"id":95,"slug":96,"title":97,"created_at":98},"ea6be18b-c903-4e54-97b7-5f7447a612e0","nvidia-gtc-2026-big-ai-announcements-zh","NVIDIA GTC 2026 重點拆解","2026-03-26T07:14:26.62638+00:00",{"id":100,"slug":101,"title":102,"created_at":103},"4bcec76f-4c36-4daa-909f-54cd702f7c93","claude-users-spreading-out-and-getting-better-zh","Claude 用戶更分散，也更會用","2026-03-26T07:22:52.325888+00:00",{"id":105,"slug":106,"title":107,"created_at":108},"bd903b15-2473-4178-9789-b7557816e535","openclaw-raises-hard-question-for-ai-models-zh","OpenClaw 逼問 AI 模型價值","2026-03-26T07:24:54.707486+00:00",{"id":110,"slug":111,"title":112,"created_at":113},"eeac6b9e-ad9d-4831-8eec-8bba3f9bca6a","gap-google-gemini-checkout-fashion-search-zh","Gap 把結帳搬進 Gemini","2026-03-26T07:28:23.937768+00:00",{"id":115,"slug":116,"title":117,"created_at":118},"0740e53f-605d-4d57-8601-c10beb126f3c","google-pushes-gemini-transition-to-march-2026-zh","Google 把 Gemini 轉換延到 2026 年 3…","2026-03-26T07:30:12.825269+00:00",{"id":120,"slug":121,"title":122,"created_at":123},"e660d801-2421-4529-8fa9-86b82b066990","metas-llama-4-benchmark-scandal-gets-worse-zh","Meta Llama 4 分數風波又擴大","2026-03-26T07:34:21.156421+00:00",{"id":125,"slug":126,"title":127,"created_at":128},"183f9e7c-e143-40bb-a6d5-67ba84a3a8bc","accenture-mistral-ai-sovereign-enterprise-deal-zh","Accenture 攜手 Mistral AI 賣主權 AI","2026-03-26T07:38:14.818906+00:00",{"id":130,"slug":131,"title":132,"created_at":133},"191d9b1b-768a-478c-978c-dd7431a38149","mistral-ai-faces-its-hardest-year-yet-zh","Mistral AI 迎來最硬的一年","2026-03-26T07:40:23.716374+00:00"]