{"id":663,"date":"2026-04-06T15:36:42","date_gmt":"2026-04-06T07:36:42","guid":{"rendered":"https:\/\/www.liaoxinghui.com\/?p=663"},"modified":"2026-04-06T15:36:42","modified_gmt":"2026-04-06T07:36:42","slug":"lora-finetune-qlora-single-gpu-7b-delivery","status":"publish","type":"post","link":"https:\/\/www.liaoxinghui.com\/?p=663","title":{"rendered":"LoRA\u5fae\u8c03\u5b9e\u6218\uff1a\u4f7f\u7528QLoRA\u5728\u5355\u5361GPU\u4e0a\u5fae\u8c0370\u4ebf\u53c2\u6570\u5927\u6a21\u578b"},"content":{"rendered":"

\u670b\u53cb\u8054\u7cfb\u6211\uff0c\u4ed6\u4eec\u662f\u4e2a\u521d\u521b\u516c\u53f8\uff0c\u60f3\u7528\u5927\u6a21\u578b\u5904\u7406\u5ba2\u670d\u5bf9\u8bdd\uff0c\u4f46\u9884\u7b97\u53ea\u591f\u4e70\u4e00\u5f204090\u663e\u5361\uff0c\u95ee\u6211\u80fd\u4e0d\u80fd\u5728\u5355\u5361\u4e0a\u5fae\u8c03\u4e2a70\u4ebf\u53c2\u6570\u7684\u6a21\u578b\u3002\u6211\u5f53\u65f6\u7b2c\u4e00\u53cd\u5e94\u662f\uff1a\u8fd9\u4e0d\u592a\u53ef\u80fd\u5427\uff1f7B\u6a21\u578b\u5168\u91cf\u5fae\u8c03\uff0c\u663e\u5b58\u81f3\u5c11\u5f9740G\u5f80\u4e0a\uff0c4090\u624d24G\u3002<\/p>\n

\u4f46\u8bdd\u4e0d\u80fd\u8bf4\u6b7b\uff0c\u6211\u6302\u4e86\u7535\u8bdd\u5c31\u5f00\u59cb\u7422\u78e8\u3002\u8fd9\u9879\u76ee\u5f97\u7528\u70b9\u5fc3\u601d\u6765\uff0c\u4e0d\u80fd\u778e\u641e\u3002<\/p>\n

\u4e1a\u52a1\u573a\u666f<\/h2>\n

\u4ed6\u4eec\u7684\u6838\u5fc3\u9700\u6c42\u5176\u5b9e\u633a\u660e\u786e\u7684\uff1a\u6709\u4e2a\u5728\u7ebf\u5546\u57ce\uff0c\u6bcf\u5929\u5927\u69825000\u6761\u5ba2\u670d\u5bf9\u8bdd\u3002\u73b0\u5728\u7684\u5ba2\u670d\u673a\u5668\u4eba\u592a\u6b7b\u677f\uff0c\u56de\u7b54\u90fd\u662f\u56fa\u5b9a\u6a21\u677f\uff0c\u670b\u53cb\u4e0d\u6ee1\u610f\u3002\u4ed6\u4eec\u60f3\u7528\u5927\u6a21\u578b\u6765\u7406\u89e3\u7528\u6237\u610f\u56fe\uff0c\u751f\u6210\u66f4\u7075\u6d3b\u3001\u66f4\u8d34\u5207\u7684\u56de\u590d\u3002<\/p>\n

\u4f46\u7ea6\u675f\u6761\u4ef6\u4e5f\u5f88\u786c\uff1a<\/p>\n

    \n
  1. \u786c\u4ef6\u53ea\u6709\u4e00\u53f0\u670d\u52a1\u5668\uff0c\u5355\u5f20NVIDIA RTX 4090\uff0824G\u663e\u5b58\uff09\u3002<\/li>\n
  2. \u6570\u636e\u4e0d\u80fd\u51fa\u516c\u53f8\uff0c\u5f97\u672c\u5730\u8bad\u7ec3\u3002<\/li>\n
  3. \u4e24\u5468\u5185\u8981\u4e0a\u7ebf\u8bd5\u8fd0\u884c\u3002<\/li>\n<\/ol>\n

    \u6211\u5f53\u65f6\u8ddf\u670b\u53cb\u5f00\u4e86\u4e2a\u4f1a\uff0c\u628a\u9700\u6c42\u5bf9\u6e05\u695a\u4e86\u3002\u9a8c\u6536\u6807\u51c6\u5c31\u4e24\u6761\uff1a\u7b2c\u4e00\uff0c\u6a21\u578b\u80fd\u57284090\u4e0a\u7a33\u5b9a\u8bad\u7ec3\uff0c\u663e\u5b58\u4e0d\u7206\uff1b\u7b2c\u4e8c\uff0c\u751f\u6210\u7684\u5ba2\u670d\u56de\u590d\u6bd4\u539f\u6765\u7684\u673a\u5668\u4eba\u5f3a\uff0c\u4eba\u5de5\u8bc4\u4f30\u901a\u8fc7\u7387\u5f97\u8d85\u8fc780%\u3002<\/p>\n

    \u67b6\u6784\u8bbe\u8ba1<\/h2>\n

    \u5b9a\u4e86\u9700\u6c42\uff0c\u63a5\u4e0b\u6765\u5c31\u662f\u6280\u672f\u9009\u578b\u3002\u5168\u91cf\u5fae\u8c03\u80af\u5b9a\u6ca1\u620f\uff0c\u663e\u5b58\u4e0d\u591f\uff0c\u90a3\u5c31LoRA\u628a\uff0c\u633a\u7b80\u5355\u7684\u3002<\/p>\n

    \u6211\u5bf9\u6bd4\u4e86\u51e0\u4e2a\u65b9\u6848\uff1a<\/p>\n

      \n
    • \u6807\u51c6LoRA\uff1a\u80fd\u7701\u663e\u5b58\uff0c\u4f46\u8bad\u7ec3\u901f\u5ea6\u6162\uff0c\u800c\u4e14\u7528\u76848bit\u91cf\u5316\uff0c\u663e\u5b58\u5360\u7528\u8fd8\u662f\u504f\u9ad8\u3002<\/li>\n
    • QLoRA\uff1a4bit\u91cf\u5316\uff0cNF4\u7c7b\u578b\uff0c\u663e\u5b58\u80fd\u518d\u780d\u4e00\u534a\uff0c\u4f46\u517c\u5bb9\u6027\u5f97\u6d4b\u8bd5\u3002<\/li>\n
    • \u5176\u4ed6\u538b\u7f29\u65b9\u6cd5\uff1a\u50cfAdapter\u3001Prefix Tuning\u4e5f\u770b\u8fc7\uff0c\u4f46\u793e\u533a\u652f\u6301\u6ca1LoRA\u597d\uff0c\u6587\u6863\u4e5f\u5c11\u3002<\/li>\n<\/ul>\n

      \u6700\u540e\u9009\u4e86QLoRA\uff0c\u4e3b\u8981\u662f\u770b\u4e2d\u5b83\u7684\u663e\u5b58\u4f18\u52bf\u3002\u57fa\u7840\u6a21\u578b\u540e\u9762\u6709\uff0c\u53c2\u6570\u5dee\u4e0d\u591a\uff0c\u80fd\u63a5\u53d7\u3002\u8bad\u7ec3\u6846\u67b6\u7528Hugging Face\u7684PEFT + bitsandbytes\uff0c\u63a8\u7406\u65f6\u518d\u628aLoRA\u6743\u91cd\u5408\u5e76\u56de\u53bb\u3002<\/p>\n

      \u5bb9\u91cf\u65b9\u9762\uff0c\u5355\u53614090\u8dd1QLoRA\uff0c\u6211\u9884\u4f30\u663e\u5b58\u5360\u7528\u80fd\u63a7\u5236\u572812G\u5de6\u53f3\uff0c\u7559\u70b9\u4f59\u91cf\u7ed9\u7cfb\u7edf\u3002\u6269\u5c55\u6027\u561b\uff0c\u8bf4\u5b9e\u8bdd\uff0c\u8fd9\u67b6\u6784\u8981\u6269\u5c55\u53ea\u80fd\u52a0\u663e\u5361\uff0c\u4f46\u9884\u7b97\u5c31\u90a3\u6837\uff0c\u5148\u8dd1\u8d77\u6765\u518d\u8bf4\u3002<\/p>\n

      \u6570\u636e\u8bf4\u660e<\/h2>\n

      \u670b\u53cb\u7ed9\u4e865000\u6761\u5ba2\u670d\u5bf9\u8bdd\u6570\u636e\uff0c\u4f46\u683c\u5f0f\u4e71\u4e03\u516b\u7cdf\u2014\u2014\u6709\u7684JSON\uff0c\u6709\u7684\u7eaf\u6587\u672c\uff0c\u8fd8\u6709\u7684\u5e26HTML\u6807\u7b7e\u3002\u6211\u62ff\u5230\u6570\u636e\u7684\u65f6\u5019\u5934\u90fd\u5927\u4e86\uff0c\u8fd9\u5f97\u6e05\u6d17\u5230\u4ec0\u4e48\u65f6\u5019\uff1f<\/p>\n

      \u6570\u636e\u6765\u6e90\u662f\u4ed6\u4eec\u7684\u5ba2\u670d\u7cfb\u7edf\u5bfc\u51fa\uff0c\u5305\u542b\u7528\u6237\u63d0\u95ee\u548c\u5ba2\u670d\u56de\u590d\u3002\u4f46\u95ee\u9898\u5f88\u591a\uff1a<\/p>\n

        \n
      1. \u91cd\u590d\u5bf9\u8bdd\u5927\u6982\u670910%\uff0c\u53ef\u80fd\u662f\u7cfb\u7edfbug\u5bfc\u81f4\u7684\u3002<\/li>\n
      2. \u6709\u4e9b\u56de\u590d\u592a\u77ed\uff0c\u5c31\u4e00\u4e24\u4e2a\u5b57\uff0c\u6bd4\u5982\u201c\u55ef\u201d\u3001\u201c\u597d\u7684\u201d\uff0c\u8fd9\u79cd\u6ca1\u8bad\u7ec3\u4ef7\u503c\u3002<\/li>\n
      3. \u683c\u5f0f\u4e0d\u7edf\u4e00\uff0c\u6709\u7684\u5b57\u6bb5\u540d\u662f\u201cquestion\u201d\uff0c\u6709\u7684\u662f\u201cquery\u201d\uff0c\u5f97\u5bf9\u9f50\u3002<\/li>\n<\/ol>\n

        \u6211\u82b1\u4e86\u5927\u69823\u4e2a\u5c0f\u65f6\u5199\u6e05\u6d17\u811a\u672c\u3002\u6838\u5fc3\u903b\u8f91\u5c31\u51e0\u6b65\uff1a<\/p>\n

        # \u6570\u636e\u6e05\u6d17\u811a\u672c\uff08\u7b80\u5316\u7248\uff09\nimport json\nimport hashlib\nfrom typing import List, Dict\n\ndef load_data(file_path: str) -> List[Dict]:\n    """\u52a0\u8f7d\u539f\u59cb\u6570\u636e\uff0c\u652f\u6301JSON\u548c\u6587\u672c\u683c\u5f0f"""\n    # \u5b9e\u9645\u4ee3\u7801\u4f1a\u6839\u636e\u4e0d\u540c\u683c\u5f0f\u505a\u89e3\u6790\uff0c\u8fd9\u91cc\u7701\u7565\n    pass\n\ndef clean_text(text: str) -> str:\n    """\u6e05\u6d17\u6587\u672c\uff1a\u53bb\u6389HTML\u6807\u7b7e\u3001\u591a\u4f59\u7a7a\u683c"""\n    import re\n    text = re.sub(r'<[^>]+>', '', text)  # \u53bbHTML\u6807\u7b7e\n    text = re.sub(r'\\s+', ' ', text).strip()  # \u5408\u5e76\u7a7a\u683c\n    return text\n\ndef deduplicate(data: List[Dict]) -> List[Dict]:\n    """\u57fa\u4e8eMD5\u53bb\u91cd"""\n    seen = set()\n    unique_data = []\n    for item in data:\n        # \u628a\u5bf9\u8bdd\u5185\u5bb9\u62fc\u63a5\u8d77\u6765\u751f\u6210MD5\n        content = item['question'] + item['answer']\n        content_hash = hashlib.md5(content.encode()).hexdigest()\n        if content_hash not in seen:\n            seen.add(content_hash)\n            unique_data.append(item)\n    return unique_data\n\ndef filter_by_length(data: List[Dict], min_len: int = 5) -> List[Dict]:\n    """\u8fc7\u6ee4\u6389\u592a\u77ed\u7684\u56de\u590d"""\n    filtered = []\n    for item in data:\n        if len(item['answer'].strip()) >= min_len:\n            filtered.append(item)\n    return filtered\n\n# \u4e3b\u6d41\u7a0b\nraw_data = load_data("customer_service_raw.json")\ncleaned_data = []\nfor item in raw_data:\n    item['question'] = clean_text(item['question'])\n    item['answer'] = clean_text(item['answer'])\n    cleaned_data.append(item)\n\ndeduped_data = deduplicate(cleaned_data)  # \u53bb\u91cd\u540e\u52694500\u6761\nfinal_data = filter_by_length(deduped_data, min_len=5)  # \u8fc7\u6ee4\u540e\u52694200\u6761\n\n# \u4fdd\u5b58\u6210\u6807\u51c6\u683c\u5f0f\nwith open("cleaned_data.json", "w", encoding="utf-8") as f:\n    json.dump(final_data, f, ensure_ascii=False, indent=2)<\/code><\/pre>\n
        \u8fd9\u91cc\u6211\u8e29\u4e86\u4e2a\u5751\uff1a\u4e00\u5f00\u59cb\u6ca1\u505a\u53bb\u91cd\uff0c\u7ed3\u679c\u8bad\u7ec3\u65f6\u6a21\u578b\u8001\u662f\u91cd\u590d\u751f\u6210\u76f8\u4f3c\u7684\u56de\u590d\u3002\u540e\u6765\u52a0\u4e86\u4e2a\u57fa\u4e8eMD5\u7684\u7b80\u5355\u53bb\u91cd\uff0c\u624d\u89e3\u51b3\u3002\u6570\u636e\u6e05\u6d17\u8fd9\u5757\uff0c\u771f\u4e0d\u80fd\u5077\u61d2\u3002<\/p>\n
        \u5b9e\u73b0\u6b65\u9aa4<\/h2>\n
        \u73af\u5883\u642d\u5efa\u6211\u7528\u7684Ubuntu 22.04\uff0cPython 3.10\uff0cPyTorch 2.1.0\u3002\u4f9d\u8d56\u5305\u4e3b\u8981\u5c31\u4e09\u4e2a\uff1atransformers\u3001peft\u3001bitsandbytes\u3002\u88c5bitsandbytes\u7684\u65f6\u5019\u6709\u70b9\u9ebb\u70e6\uff0c\u5f97\u4ece\u6e90\u7801\u7f16\u8bd1\uff0c\u6211\u6298\u817e\u4e86\u5feb\u4e00\u4e2a\u5c0f\u65f6\u624d\u641e\u5b9a\u3002<\/p>\n
        \u6a21\u578b\u52a0\u8f7d\u7684\u4ee3\u7801\u662f\u5173\u952e\uff0cQLoRA\u7684\u914d\u7f6e\u90fd\u5728\u8fd9\u91cc\uff1a<\/p>\n
        # \u8fd9\u662f\u6700\u7ec8\u53ef\u8fd0\u884c\u7684\u6a21\u578b\u52a0\u8f7d\u4ee3\u7801\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nfrom peft import LoraConfig, get_peft_model\nimport torch\nfrom transformers import BitsAndBytesConfig\n\n# \u5173\u952e\u914d\u7f6e\uff1a4bit\u91cf\u5316\uff0c\u7528NF4\u7c7b\u578b\nbnb_config = BitsAndBytesConfig(\n    load_in_4bit=True,\n    bnb_4bit_quant_type="nf4",  # \u4e00\u5b9a\u8981\u7528NF4\uff0c\u522b\u7528FP4\n    bnb_4bit_compute_dtype=torch.float16,  # \u8ba1\u7b97\u7528float16\u52a0\u901f\n    bnb_4bit_use_double_quant=True  # \u53cc\u5c42\u91cf\u5316\uff0c\u518d\u7701\u70b9\u5185\u5b58\n)\n\n# \u52a0\u8f7d\u57fa\u7840\u6a21\u578b\nmodel = AutoModelForCausalLM.from_pretrained(\n    "THUDM\/chatglm2-6b",\n    quantization_config=bnb_config,\n    device_map="auto",  # \u81ea\u52a8\u5206\u914d\u8bbe\u5907\n    trust_remote_code=True\n)\ntokenizer = AutoTokenizer.from_pretrained("THUDM\/chatglm2-6b", trust_remote_code=True)\n\n# LoRA\u914d\u7f6e\uff1a\u53ea\u8bad\u7ec3attention\u7684qkv\u548c\u8f93\u51fa\u6295\u5f71\nlora_config = LoraConfig(\n    r=8,  # \u79e9\uff0c\u6211\u8bd5\u8fc74\u30018\u300116\uff0c8\u5728\u8fd9\u4e2a\u4efb\u52a1\u4e0a\u6548\u679c\u6700\u597d\n    lora_alpha=32,\n    target_modules=["query_key_value", "dense"],  # ChatGLM2\u7684\u7279\u6b8a\u7ed3\u6784\n    lora_dropout=0.1,\n    bias="none",\n    task_type="CAUSAL_LM"\n)\nmodel = get_peft_model(model, lora_config)\nmodel.print_trainable_parameters()  # \u8f93\u51fa\uff1atrainable params: 3,932,160 || all params: 6,248,320,000 || trainable%: 0.0629%<\/code><\/pre>\n
        \u8dd1\u8d77\u6765\u540e\uff0c\u663e\u5b58\u5360\u7528\u5927\u698212G\uff0c4090\u5b8c\u5168\u625b\u5f97\u4f4f\u3002\u4f46\u8bad\u7ec3\u53c2\u6570\u8c03\u4f18\u53c8\u662f\u4e2a\u5934\u75bc\u7684\u4e8b\u3002<\/p>\n
        \u53c2\u6570\u8bf4\u660e<\/h2>\n
        \u6211\u4e00\u5f00\u59cb\u5b66\u4e60\u7387\u8bbe\u4e862e-4\uff0c\u7ed3\u679closs\u9707\u8361\u5f97\u5389\u5bb3\uff0c\u50cf\u5750\u8fc7\u5c71\u8f66\u3002\u540e\u6765\u8c03\u52305e-5\u624d\u7a33\u5b9a\u4e0b\u6765\u3002\u8fd9\u91cc\u6709\u4e2a\u7ecf\u9a8c\uff1aQLoRA\u56e0\u4e3a\u53c2\u6570\u5c11\uff0c\u5b66\u4e60\u7387\u8981\u6bd4\u5168\u91cf\u5fae\u8c03\u8bbe\u5c0f\u4e00\u70b9\u3002<\/p>\n\n\n\n\n\n\n\n\n\n\n
        \u53c2\u6570<\/th>\n\u63a8\u8350\u503c<\/th>\n\u6211\u8bd5\u8fc7\u7684\u8303\u56f4<\/th>\n\u8bf4\u660e<\/th>\n<\/tr>\n<\/thead>\n
        \u5b66\u4e60\u7387<\/td>\n5e-5<\/td>\n1e-4 ~ 1e-5<\/td>\n\u5927\u4e86\u5bb9\u6613\u9707\u8361\uff0c\u5c0f\u4e86\u6536\u655b\u6162<\/td>\n<\/tr>\n
        \u6279\u5927\u5c0f<\/td>\n8<\/td>\n4 ~ 16<\/td>\n\u53d7\u663e\u5b58\u9650\u5236\uff0c\u6211\u6700\u5927\u53ea\u80fd\u52308<\/td>\n<\/tr>\n
        \u8bad\u7ec3\u8f6e\u6570<\/td>\n3<\/td>\n1 ~ 5<\/td>\n\u591a\u4e86\u5bb9\u6613\u8fc7\u62df\u5408<\/td>\n<\/tr>\n
        LoRA\u79e9(r)<\/td>\n8<\/td>\n4, 8, 16<\/td>\n8\u6027\u4ef7\u6bd4\u6700\u9ad8<\/td>\n<\/tr>\n
        \u4f18\u5316\u5668<\/td>\nAdamW<\/td>\n–<\/td>\n\u9ed8\u8ba4\u7684\u5c31\u884c<\/td>\n<\/tr>\n
        \u5b66\u4e60\u7387\u8c03\u5ea6<\/td>\ncosine<\/td>\n–<\/td>\n\u6162\u6162\u4e0b\u964d\uff0c\u907f\u514d\u7a81\u53d8<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
        \u8bad\u7ec3\u4e863\u4e2aepoch\uff0closs\u66f2\u7ebf\u662f\u8fd9\u6837\u7684\uff1a<\/p>\n
        Epoch 1: loss 2.34 -> 1.87\nEpoch 2: loss 1.87 -> 1.45\nEpoch 3: loss 1.45 -> 1.23<\/code><\/pre>\n
        \u4e0b\u964d\u8d8b\u52bf\u8fd8\u884c\uff0c\u4f46\u5230\u7b2c3\u8f6e\u5df2\u7ecf\u6709\u70b9\u5e73\u4e86\uff0c\u518d\u8bad\u7ec3\u53ef\u80fd\u5c31\u8fc7\u62df\u5408\u4e86\u3002\u6211\u5f53\u65f6\u505c\u5728\u4e86\u7b2c3\u8f6e\uff0c\u4fdd\u5b58\u4e86checkpoint\u3002<\/p>\n
        \u8c03\u7528\u65b9\u5f0f<\/h2>\n
        \u8bad\u7ec3\u5b8c\u7684\u6a21\u578b\u4e0d\u80fd\u76f4\u63a5\u62ff\u6765\u7528\uff0c\u5f97\u628aLoRA\u6743\u91cd\u5408\u5e76\u56de\u539f\u6a21\u578b\uff0c\u4e0d\u7136\u52a0\u8f7d\u6162\uff0c\u63a8\u7406\u4e5f\u9ebb\u70e6\u3002\u6211\u7528\u7684\u662fPEFT\u7684merge_and_unload\u65b9\u6cd5\u3002<\/p>\n
        # \u63a8\u7406\u793a\u4f8b\uff1a\u5408\u5e76\u6743\u91cd\u5e76\u4fdd\u5b58\nfrom peft import PeftModel\n\n# \u52a0\u8f7d\u57fa\u7840\u6a21\u578b\uff08\u8ddf\u8bad\u7ec3\u65f6\u4e00\u6837\u7684\u914d\u7f6e\uff09\nbase_model = AutoModelForCausalLM.from_pretrained(\n    "THUDM\/chatglm2-6b",\n    quantization_config=bnb_config,\n    device_map="auto",\n    trust_remote_code=True\n)\n\n# \u52a0\u8f7dLoRA\u6743\u91cd\nmodel = PeftModel.from_pretrained(base_model, ".\/lora_checkpoint")\n\n# \u5408\u5e76\u6743\u91cd\nmodel = model.merge_and_unload()\n\n# \u4fdd\u5b58\u5408\u5e76\u540e\u7684\u6a21\u578b\nmodel.save_pretrained(".\/merged_model")\ntokenizer.save_pretrained(".\/merged_model")\n\n# \u5b9e\u9645\u63a8\u7406\u4ee3\u7801\ninput_text = "\u7528\u6237\u95ee\uff1a\u9000\u8d27\u600e\u4e48\u5904\u7406\uff1f"\ninputs = tokenizer(input_text, return_tensors="pt").to(model.device)\noutputs = model.generate(**inputs, max_length=100, temperature=0.8)\nresponse = tokenizer.decode(outputs[0], skip_special_tokens=True)\nprint(response)<\/code><\/pre>\n
        \u5408\u5e76\u540e\u7684\u6a21\u578b\u5c31\u8ddf\u666e\u901a\u6a21\u578b\u4e00\u6837\u7528\u4e86\uff0c\u90e8\u7f72\u8d77\u6765\u4e5f\u7b80\u5355\u3002\u6211\u5efa\u8bae\u4fdd\u5b58\u4e24\u4e2a\u7248\u672c\uff1a\u4e00\u4e2a\u662f\u5e26LoRA\u6743\u91cd\u7684\uff08\u65b9\u4fbf\u7ee7\u7eed\u8bad\u7ec3\uff09\uff0c\u4e00\u4e2a\u662f\u5408\u5e76\u540e\u7684\uff08\u7528\u4e8e\u63a8\u7406\uff09\u3002<\/p>\n
        \u4e0a\u7ebf\u540e\u8bc4\u4f30<\/h2>\n
        \u6a21\u578b\u8bad\u7ec3\u5b8c\uff0c\u5f97\u9a8c\u6536\u4e86\u3002\u6211\u6ca1\u7528\u90a3\u4e9b\u590d\u6742\u7684BLEU\u3001ROUGE\u6307\u6807\u2014\u2014\u8bf4\u5b9e\u8bdd\uff0c\u5728\u5ba2\u670d\u5bf9\u8bdd\u573a\u666f\uff0c\u90a3\u4e9b\u6307\u6807\u8ddf\u4eba\u5de5\u8bc4\u4ef7\u5dee\u8ddd\u633a\u5927\u7684\u3002\u6211\u76f4\u63a5\u627e\u4e8610\u4e2a\u5178\u578b\u95ee\u9898\uff0c\u8ba9\u5fae\u8c03\u524d\u540e\u7684\u6a21\u578b\u90fd\u56de\u7b54\uff0c\u4eba\u5de5\u5bf9\u6bd4\u3002<\/p>\n
        \u6d4b\u8bd5\u7528\u4f8b\u6bd4\u5982\uff1a<\/p>\n
          \n
        1. \u201c\u9000\u8d27\u600e\u4e48\u5904\u7406\uff1f\u201d<\/li>\n
        2. \u201c\u8ba2\u5355\u4e00\u76f4\u6ca1\u53d1\u8d27\uff0c\u600e\u4e48\u56de\u4e8b\uff1f\u201d<\/li>\n
        3. \u201c\u5546\u54c1\u6709\u7834\u635f\uff0c\u80fd\u6362\u8d27\u5417\uff1f\u201d<\/li>\n<\/ol>\n
          \u539f\u59cb\u6a21\u578b\u7684\u56de\u7b54\u6bd4\u8f83\u901a\u7528\uff0c\u6bd4\u5982\u201c\u8bf7\u63d0\u4f9b\u8ba2\u5355\u53f7\uff0c\u6211\u4eec\u5c06\u4e3a\u60a8\u5904\u7406\u9000\u8d27\u201d\u3002\u800c\u5fae\u8c03\u540e\u7684\u6a21\u578b\u4f1a\u5e26\u4e0a\u670b\u53cb\u5177\u4f53\u7684\u653f\u7b56\uff0c\u6bd4\u5982\u201c\u6839\u636e\u9000\u8d27\u653f\u7b56P-2023-001\uff0c\u8bf7\u5728\u6536\u5230\u5546\u54c17\u5929\u5185\u7533\u8bf7\u9000\u8d27\uff0c\u5e76\u786e\u4fdd\u5546\u54c1\u5b8c\u597d\u201d\u3002<\/p>\n
          \u6211\u4eec\u56e2\u961f3\u4e2a\u4eba\u4e00\u8d77\u8bc4\u4f30\uff0c\u6bcf\u4eba\u7ed9\u56de\u590d\u6253\u5206\uff081-5\u5206\uff09\uff0c\u6700\u540e\u7b97\u5e73\u5747\u5206\u3002\u5fae\u8c03\u524d\u6a21\u578b\u5e73\u57472.8\u5206\uff0c\u5fae\u8c03\u540e\u5e73\u57474.2\u5206\uff0c\u901a\u8fc7\u7387\uff084\u5206\u4ee5\u4e0a\uff0985%\uff0c\u8d85\u8fc7\u4e86\u670b\u53cb\u8981\u6c42\u768480%\u3002<\/p>\n
          \u4e0a\u7ebf\u540e\u76d1\u63a7\u4e5f\u5f88\u91cd\u8981\u3002\u6211\u8ba9\u670b\u53cb\u5728\u5ba2\u670d\u7cfb\u7edf\u91cc\u52a0\u4e86\u4e2a\u53cd\u9988\u6309\u94ae\uff0c\u7528\u6237\u53ef\u4ee5\u5bf9\u673a\u5668\u4eba\u56de\u590d\u6253\u5206\u3002\u5934\u4e00\u5468\u7684\u6570\u636e\u663e\u793a\uff0c\u6ee1\u610f\u5ea6\u5927\u6982\u572870%\u5de6\u53f3\uff0c\u6bd4\u4e4b\u524d\u768450%\u5f3a\u591a\u4e86\u3002\u5f53\u7136\uff0c\u8fd8\u6709\u4f18\u5316\u7a7a\u95f4\uff0c\u4f46\u7b2c\u4e00\u671f\u4ea4\u4ed8\u7b97\u8fbe\u6807\u4e86\u3002<\/p>\n
          \u5e38\u89c1\u5751<\/h2>\n
          \u6574\u4e2a\u9879\u76ee\u505a\u4e0b\u6765\uff0c\u8e29\u7684\u5751\u4e0d\u5c11\uff0c\u6211\u603b\u7ed3\u51e0\u4e2a\u4e3b\u8981\u7684\uff1a<\/p>\n
            \n
          1. \n
            \u91cf\u5316\u7c7b\u578b\u9009\u9519<\/strong>\uff1a\u4e00\u5f00\u59cb\u7528\u4e86FP4\uff0c\u6548\u679c\u6bd4NF4\u5dee\u4e00\u622a\uff0closs\u4e0b\u4e0d\u53bb\u3002\u540e\u6765\u67e5\u6587\u6863\u624d\u53d1\u73b0\uff0cNF4\u662f\u4e13\u95e8\u4e3a\u795e\u7ecf\u7f51\u7edc\u4f18\u5316\u7684\uff0cFP4\u662f\u901a\u7528\u91cf\u5316\u3002<\/p>\n<\/li>\n
          2. \n
            \u6570\u636e\u6e05\u6d17\u5077\u61d2<\/strong>\uff1a\u6ca1\u53bb\u91cd\u7684\u90a3\u7248\u6a21\u578b\uff0c\u751f\u6210\u56de\u590d\u91cd\u590d\u7387\u9ad8\u8fbe30%\uff0c\u6839\u672c\u6ca1\u6cd5\u7528\u3002\u6570\u636e\u8d28\u91cf\u592a\u5173\u952e\u4e86\u3002<\/p>\n<\/li>\n
          3. \n
            \u5b66\u4e60\u7387\u8bbe\u592a\u9ad8<\/strong>\uff1aQLoRA\u53c2\u6570\u5c11\uff0c2e-4\u7684\u5b66\u4e60\u7387\u76f4\u63a5\u8bad\u98de\u4e86\uff0closs\u4e0a\u8e7f\u4e0b\u8df3\u3002\u8c03\u52305e-5\u624d\u7a33\u4e0b\u6765\u3002<\/p>\n<\/li>\n
          4. \n
            \u8bc4\u4f30\u65b9\u6cd5\u8131\u79bb\u4e1a\u52a1<\/strong>\uff1a\u5149\u770bloss\u4e0b\u964d\u6ca1\u7528\uff0c\u5f97\u7ed3\u5408\u5b9e\u9645\u573a\u666f\u6d4b\u8bd5\u3002\u4eba\u5de5\u8bc4\u4f30\u867d\u7136\u8d39\u65f6\uff0c\u4f46\u6700\u9760\u8c31\u3002<\/p>\n<\/li>\n
          5. \n
            \u5fd8\u8bb0\u5408\u5e76\u6743\u91cd<\/strong>\uff1a\u7b2c\u4e00\u6b21\u90e8\u7f72\u65f6\u76f4\u63a5\u52a0\u8f7d\u4e86LoRA\u6743\u91cd\uff0c\u63a8\u7406\u901f\u5ea6\u6162\u4e0d\u8bf4\uff0c\u8fd8\u591a\u5360\u663e\u5b58\u3002\u540e\u6765\u624d\u60f3\u8d77\u6765\u8981merge_and_unload\u3002<\/p>\n<\/li>\n<\/ol>\n
            \u6f14\u8fdb\u8def\u7ebf<\/h2>\n
            \u8fd9\u9879\u76ee\u76ee\u524d\u7b97\u662f\u4ea4\u4ed8\u4e86\uff0c\u4f46\u957f\u8fdc\u770b\u8fd8\u6709\u4f18\u5316\u7a7a\u95f4\u3002\u6211\u8ddf\u670b\u53cb\u804a\u8fc7\u540e\u7eed\u7684\u6f14\u8fdb\u8def\u7ebf\uff1a<\/p>\n
              \n
            1. \u77ed\u671f\uff081\u4e2a\u6708\uff09<\/strong>\uff1a\u6536\u96c6\u66f4\u591a\u7528\u6237\u53cd\u9988\u6570\u636e\uff0c\u505a\u7b2c\u4e8c\u8f6e\u5fae\u8c03\uff0c\u91cd\u70b9\u4f18\u5316\u90a3\u4e9b\u5f97\u5206\u4f4e\u7684case\u3002<\/li>\n
            2. \u4e2d\u671f\uff083\u4e2a\u6708\uff09<\/strong>\uff1a\u5982\u679c\u4e1a\u52a1\u91cf\u589e\u957f\uff0c\u8003\u8651\u5347\u7ea7\u663e\u5361\uff0c\u6216\u8005\u7528\u591a\u5361\u5e76\u884c\u8bad\u7ec3\u66f4\u5927\u7684\u6a21\u578b\u3002<\/li>\n
            3. \u957f\u671f\uff086\u4e2a\u6708\uff09<\/strong>\uff1a\u63a2\u7d22\u5168\u91cf\u5fae\u8c03\u6216\u8005\u66f4\u9ad8\u7ea7\u7684\u5fae\u8c03\u65b9\u6cd5\uff0c\u6bd4\u5982DoRA\uff0c\u4f46\u5f97\u770b\u9884\u7b97\u60c5\u51b5\u3002<\/li>\n<\/ol>\n
              \u6574\u4e2a\u9879\u76ee\u4ece\u9700\u6c42\u5bf9\u63a5\u5230\u4e0a\u7ebf\u8bc4\u4f30\uff0c\u82b1\u4e86\u5927\u6982\u4e24\u5468\u65f6\u95f4\u3002\u73b0\u5728\u6a21\u578b\u5df2\u7ecf\u5728\u670b\u53cb\u90a3\u8fb9\u8dd1\u8d77\u6765\u4e86\uff0c\u8d44\u6e90\u5360\u7528\u6bd4\u9884\u671f\u4f4e\uff0c\u6548\u679c\u4e5f\u8fd8\u884c\u3002\u5f53\u7136\uff0c\u5982\u679c\u663e\u5361\u518d\u591a\u70b9\uff0c\u6211\u53ef\u80fd\u4f1a\u8bd5\u8bd5\u66f4\u5927\u7684\u79e9\u6216\u8005\u5168\u91cf\u5fae\u8c03\uff0c\u4f46\u5355\u5361\u73af\u5883\u4e0b\uff0cQLoRA\u786e\u5b9e\u662f\u76ee\u524d\u6700\u5b9e\u7528\u7684\u65b9\u6848\u4e86\u3002\u505a\u9879\u76ee\u4ea4\u4ed8\uff0c\u5173\u952e\u662f\u628a\u8fb9\u754c\u753b\u6e05\u695a\uff0c\u6bcf\u4e2a\u9636\u6bb5\u90fd\u6709\u660e\u786e\u7684\u4ea7\u51fa\uff0c\u8fd9\u6837\u624d\u4e0d\u5bb9\u6613\u7ffb\u8f66\u3002<\/p>","protected":false},"excerpt":{"rendered":"
              \u670b\u53cb\u60f3\u5728\u5355\u5f204090\u4e0a\u5fae\u8c037B\u6a21\u578b\u505a\u5ba2\u670d\uff0c\u6211\u4e00\u5f00\u59cb\u89c9\u5f97\u4e0d\u53ef\u80fd\u3002\u6309\u9879\u76ee\u4ea4\u4ed8\u624b\u518c\u8d70\uff0c\u4ece\u4e1a\u52a1\u573a\u666f\u3001\u6570\u636e\u6e05\u6d17\u5230\u4e0a\u7ebf\u8bc4\u4f30\uff0c\u8e29\u4e86\u4e00\u5806\u5751\uff0c\u6700\u540e\u7528QLoRA\u641e\u5b9a\u4e86\u3002\u8fd9\u91cc\u5206\u4eab\u4e0b\u5b8c\u6574\u7684\u4ea4\u4ed8\u8fc7\u7a0b\u548c\u53ef\u8fd0\u884c\u7684\u811a\u672c\u3002<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[401],"tags":[402,320,403,319,382,293],"class_list":["post-663","post","type-post","status-publish","format-standard","hentry","category-ai","tag-lora","tag-qlora","tag-gpu","tag-319","tag-382","tag-293"],"views":4,"_links":{"self":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/663","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=663"}],"version-history":[{"count":1,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/663\/revisions"}],"predecessor-version":[{"id":675,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=\/wp\/v2\/posts\/663\/revisions\/675"}],"wp:attachment":[{"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=663"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=663"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.liaoxinghui.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=663"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}