{"id":115055,"date":"2026-10-11T01:57:08","date_gmt":"2026-10-10T17:57:08","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/115055.html"},"modified":"2026-10-11T01:57:08","modified_gmt":"2026-10-10T17:57:08","slug":"paged-optimizers-%e6%98%be%e5%ad%98%e6%8d%a2%e5%85%a5%e5%bb%b6%e8%bf%9f%e6%b2%bb%e7%90%86%ef%bc%9aqlora-%e8%ae%ad%e7%bb%83%e5%90%9e%e5%90%90%e8%b0%83%e4%bc%98%e5%ae%9e%e6%88%98","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/115055.html","title":{"rendered":"Paged Optimizers \u663e\u5b58\u6362\u5165\u5ef6\u8fdf\u6cbb\u7406\uff1aQLoRA \u8bad\u7ec3\u541e\u5410\u8c03\u4f18\u5b9e\u6218"},"content":{"rendered":"

\"\u5c01\u9762\u4fe1\u606f\u56fe\"<\/p>\n

\u5728\u6d88\u8d39\u7ea7\u5355\u5361（\u5982 RTX 4090 24GB）\u6216\u5355\u5f20 A100 \u4e0a\u5fae\u8c03 70B \u7ea7\u522b\u5927\u6a21\u578b\u65f6，**QLoRA + \u5206\u9875\u4f18\u5316\u5668（Paged Optimizers）**\u88ab\u8a89\u4e3a\u7a77\u4eba\u7b97\u529b\u7684\u6551\u4e16\u4e3b\u3002\u901a\u8fc7 bitsandbytes \u63d0\u4f9b\u7684\u5206\u9875\u673a\u5236，\u5f53\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6fc0\u6d3b\u503c\u6216\u68af\u5ea6\u53d1\u751f\u77ac\u65f6\u5c16\u523a\u5bfc\u81f4\u663e\u5b58\u5373\u5c06\u6ea2\u51fa（OOM）\u65f6，\u5e95\u5c42\u9a71\u52a8\u4f1a\u81ea\u52a8\u5c06\u4f18\u5316\u5668\u72b6\u6001\u9875（Optimizer States）\u5e73\u6ed1\u9010\u51fa\u5230\u7cfb\u7edf\u4e3b\u673a\u5185\u5b58（CPU RAM）\u4e2d，\u5f85\u663e\u5b58\u56de\u843d\u540e\u518d\u52a8\u6001\u6362\u5165\u3002<\/p>\n

\u8fd9\u786e\u5b9e\u8ba9\u5f88\u591a\u539f\u672c\u201c\u8dd1\u4e0d\u8d77\u6765\u201d\u7684\u4efb\u52a1\u6210\u529f\u8dd1\u901a\u4e86\u3002\u7136\u800c，\u5728\u8bb8\u591a\u56e2\u961f\u7684\u5b9e\u9645\u4f53\u9a8c\u4e2d，\u968f\u4e4b\u800c\u6765\u7684\u5374\u662f\u4e00\u573a\u6027\u80fd\u5669\u68a6：\u8bad\u7ec3\u901f\u5ea6\u5947\u6162\u65e0\u6bd4\u3002<\/p>\n

\u539f\u672c\u9884\u8ba1\u4e00\u5929\u8dd1\u5b8c\u7684\u5fae\u8c03\u4efb\u52a1，\u53ef\u80fd\u88ab\u62d6\u5ef6\u5230\u6574\u6574\u4e00\u5468\u3002\u89c2\u5bdf nvidia-smi，GPU \u7684\u7b97\u529b\u5229\u7528\u7387（SM Active）\u5e38\u5e74\u5904\u4e8e 20%~30% \u7684\u4f4e\u8c37，GPU \u4eff\u4f5b\u5728\u5927\u90e8\u5206\u65f6\u95f4\u91cc\u5904\u4e8e\u6df1\u5ea6\u7761\u7720\u3002<\/p>\n

\u5256\u6790\u5e95\u5c42\u7684\u7cfb\u7edf\u603b\u7ebf\u8c03\u7528\u540e，\u95ee\u9898\u4e00\u76ee\u4e86\u7136：GPU \u5e76\u4e0d\u662f\u5728\u7b97\u77e9\u9635，\u800c\u662f\u5728\u6f2b\u957f\u5730\u7b49\u5f85\u4f18\u5316\u5668\u72b6\u6001\u5728 PCIe \u6162\u901f\u901a\u9053\u4e2d\u6765\u56de\u642c\u8fd0\u3002\u8981\u8ba9 QLoRA \u771f\u6b63\u5177\u5907\u5de5\u4e1a\u751f\u4ea7\u7ea7\u7684\u541e\u5410\u901f\u5ea6，\u5fc5\u987b\u5bf9\u5206\u9875\u4f18\u5316\u5668\u7684\u6362\u5165\u6362\u51fa\u673a\u5236\u5b9e\u65bd\u7cfb\u7edf\u6027\u7684\u5de5\u7a0b\u6cbb\u7406\u3002<\/p>\n

PCIe \u74f6\u9888\u4e0e\u5206\u9875\u72b6\u6001\u98a0\u7c38\u65f6\u5e8f：
\n[GPU \u663e\u5b58\u6c60 (HBM \u5e26\u5bbd 2000 GB\/s)]
\n \u2502
\n \u25b2 \u53d1\u751f\u77ac\u65f6\u663e\u5b58\u5c16\u523a (Memory Spike)
\n \u25bc
\n[PCIe 4.0\/5.0 \u72ed\u7a84\u901a\u9053 (\u7269\u7406\u5e26\u5bbd\u4ec5 32~64 GB\/s，\u4e25\u91cd\u74f6\u9888)]
\n \u2502
\n \u25bc \u9a71\u52a8\u7ea7\u5206\u9875\u9a71\u9010 (Page-out to RAM)
\n[CPU \u4e3b\u673a\u5185\u5b58 (DDR5 RAM)]
\n(\u53cd\u5411\u4f20\u64ad\u66f4\u65b0\u53c2\u6570\u65f6，\u5fc5\u987b\u540c\u6b65\u7b49\u5f85\u5343\u5146\u5b57\u8282\u72b6\u6001\u91cd\u65b0\u704c\u5165\u663e\u5b58，\u8ba1\u7b97\u6838\u5fc3\u5168\u9762\u505c\u5de5\u6302\u8d77)<\/p>\n

\u4e00\u3001\u5206\u9875\u673a\u5236\u5f15\u53d1\u6027\u80fd\u96ea\u5d29\u7684\u7269\u7406\u6df1\u6e0a<\/h3>\n

\u4e3a\u4ec0\u4e48\u5206\u9875\u4f18\u5316\u5668\u4f1a\u5bfc\u81f4\u541e\u5410\u66b4\u8dcc？\u6838\u5fc3\u77db\u76fe\u5728\u4e8e\u8ba1\u7b97\u4f53\u7cfb\u7ed3\u6784\u4e2d\u7684\u5e26\u5bbd\u9e3f\u6c9f（Bandwidth Chasm）：<\/p>\n

  • \u9ad8\u8fbe 40 \u500d\u7684\u5e26\u5bbd\u65ad\u5d16：\u9ad8\u7aef GPU（\u5982 H100\/A100）\u7684\u7247\u4e0a HBM \u663e\u5b58\u5e26\u5bbd\u666e\u904d\u5728 1.5TB\/s \u5230 3.3TB\/s \u4e4b\u95f4；\u800c\u5728\u6d88\u8d39\u7ea7\u6216\u6807\u51c6\u670d\u52a1\u5668\u4e0a，PCIe 4.0 x16 \u7684\u5355\u5411\u7406\u8bba\u5e26\u5bbd\u4ec5\u4e3a 31.5GB\/s（\u5373\u4fbf\u662f PCIe 5.0 \u4e5f\u4ec5\u4e3a 63GB\/s）\u3002\u5f53\u6570\u5341\u4ebf\u53c2\u6570\u7684\u4f18\u5316\u5668\u4e00\u9636\u4e0e\u4e8c\u9636\u52a8\u91cf\u5728\u663e\u5b58\u548c\u4e3b\u673a\u5185\u5b58\u4e4b\u95f4\u9891\u7e41\u7a7f\u68ad\u65f6，PCIe \u603b\u7ebf\u77ac\u95f4\u88ab\u6253\u5230\u9971\u548c，\u7b97\u529b\u6838\u5fc3\u53ea\u80fd\u88ab\u52a8\u81ea\u65cb\u7b49\u5f85；<\/li>\n
  • \u540c\u6b65\u963b\u585e\u5f0f\u9875\u9762\u9519\u8bef（Page Fault Stall）：CUDA \u9a71\u52a8\u5c42\u9762\u7684\u7edf\u4e00\u5185\u5b58（Unified Memory）\u5206\u9875\u901a\u5e38\u662f\u6309\u9700\u540c\u6b65\u89e6\u53d1\u7684\u3002\u5f53 AdamW \u5728\u8ba1\u7b97 $W_{t+1} = W_t – \\\\eta \\\\frac{m_t}{\\\\sqrt{v_t} + \\\\epsilon}$ \u65f6，\u53ea\u8981\u8bbf\u95ee\u5230\u4e00\u4e2a\u4f4d\u4e8e CPU \u7684\u5185\u5b58\u9875，\u6574\u4e2a\u6d41\u591a\u5904\u7406\u5668（SM）\u7684\u6267\u884c\u6d41\u6c34\u7ebf\u5c31\u4f1a\u88ab\u786c\u6027\u6253\u65ad，\u7b49\u5f85 DMA \u642c\u8fd0\u5b8c\u6bd5，\u9020\u6210\u6781\u5176\u788e\u7247\u5316\u7684\u5185\u6838\u6c14\u6ce1；<\/li>\n
  • \u5185\u5b58\u98a0\u7c38（Thrashing）：\u82e5\u6bcf\u4e2a\u5fae\u6279\u6b21（Micro-batch）\u7684\u663e\u5b58\u5360\u7528\u6070\u597d\u5361\u5728\u663e\u5b58\u5bb9\u91cf\u7684\u4e34\u754c\u7ea2\u7ebf\u4e0a，\u524d\u5411\u4f20\u64ad\u521a\u628a\u4f18\u5316\u5668\u8d76\u5230 CPU，\u53cd\u5411\u4f20\u64ad\u53c8\u628a\u5b83\u62c9\u56de GPU，\u4e0b\u4e00\u8f6e\u8fed\u4ee3\u53c8\u91cd\u590d\u8fd9\u4e00\u8fc7\u7a0b，\u7cfb\u7edf\u5f7b\u5e95\u9677\u5165\u4e86\u201c\u53ea\u642c\u5bb6\u3001\u4e0d\u5e72\u6d3b\u201d\u7684\u98a0\u7c38\u6b7b\u5faa\u73af\u3002<\/li>\n

    \u4e8c\u3001\u6cbb\u7406\u4e09\u5927\u7ec4\u5408\u62f3：\u8ba9\u5206\u9875\u6210\u4e3a\u201c\u5b89\u5168\u515c\u5e95\u201d\u800c\u975e\u201c\u5e38\u6001\u901a\u8def\u201d<\/h3>\n

    \u8981\u6062\u590d GPU \u7684\u5168\u901f\u72c2\u98d9，\u6cbb\u7406\u54f2\u5b66\u662f\u660e\u786e\u7684：\u901a\u8fc7\u7cbe\u5bc6\u7684\u663e\u5b58\u9884\u7b97\u89c4\u5212，\u8ba9\u4f18\u5316\u5668\u72b6\u6001\u5e38\u9a7b GPU \u663e\u5b58；\u5206\u9875\u673a\u5236\u4ec5\u4f5c\u4e3a\u5e94\u5bf9\u7f55\u89c1\u8d85\u957f\u5e8f\u5217\u7684\u6700\u540e\u4e00\u9053\u9632\u7206\u5b89\u5168\u5e26，\u575a\u51b3\u675c\u7edd\u5e38\u6001\u5316\u6362\u5165\u6362\u51fa\u3002<\/p>\n

    1. \u91c7\u7528 8-bit Paged AdamW \u538b\u51cf 75% \u72b6\u6001\u4f53\u79ef<\/h4>\n

    \u6807\u51c6\u7684 32-bit AdamW \u4e3a\u6bcf\u4e2a\u53c2\u6570\u4fdd\u5b58\u4e24\u4e2a FP32 \u72b6\u6001（\u4e00\u9636\u52a8\u91cf\u4e0e\u4e8c\u9636\u52a8\u91cf），\u6bcf\u4e2a\u53c2\u6570\u6d88\u8017 8 \u5b57\u8282；\u800c\u91c7\u7528\u975e\u7ebf\u6027\u91cf\u5316\u7684 paged_adamw_8bit，\u6bcf\u4e2a\u52a8\u91cf\u4ec5\u5360\u7528 1 \u5b57\u8282（\u603b\u8ba1 2 \u5b57\u8282），\u76f4\u63a5\u5c06\u6f5c\u5728\u7684\u642c\u8fd0\u6570\u636e\u91cf\u538b\u7f29\u4e86 75%，\u5927\u5e45\u524a\u51cf\u4e86\u603b\u7ebf\u4f20\u8f93\u8d1f\u62c5\u3002<\/p>\n

    2. \u52a8\u6001\u68af\u5ea6\u7d2f\u52a0\u4e0e\u5fae\u6279\u6b21\u538b\u5e73（Micro-batch Flattener）<\/h4>\n

    \u5bfc\u81f4\u663e\u5b58\u53d1\u751f\u5c16\u523a\u7684\u4e3b\u8981\u6765\u6e90\u662f\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u4e2d\u95f4\u6fc0\u6d3b\u503c\u3002\u5c06\u5355\u6b65\u6279\u6b21\u5927\u5c0f（per_device_train_batch_size）\u538b\u51cf\u81f3 1 \u6216 2，\u540c\u65f6\u6210\u500d\u653e\u5927\u68af\u5ea6\u7d2f\u52a0\u6b65\u6570（gradient_accumulation_steps），\u80fd\u591f\u5c06\u524d\u5411\u6fc0\u6d3b\u663e\u5b58\u538b\u5230\u6781\u4f4e，\u4ece\u800c\u4e3a 8-bit \u4f18\u5316\u5668\u7559\u51fa\u5145\u8db3\u7684 GPU \u9a7b\u7559\u7a7a\u95f4\u3002<\/p>\n

    \u4ee5\u4e0b\u662f\u5b9e\u73b0\u9ad8\u541e\u5410 QLoRA \u8bad\u7ec3\u914d\u7f6e\u4e0e\u5206\u9875\u5f00\u9500\u76d1\u63a7\u7684\u6838\u5fc3\u4ee3\u7801：<\/p>\n

    import torch
    \nfrom transformers import TrainingArguments, Trainer
    \nfrom peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
    \nfrom transformers import AutoModelForCausalLM, BitsAndBytesConfig<\/p>\n

    def setup_high_throughput_qlora(model_name: str, output_dir: str) -> Trainer:
    \n # 1. 4-bit \u6fc0\u8fdb\u53cc\u91cf\u5316\u914d\u7f6e
    \n bnb_config = BitsAndBytesConfig(
    \n load_in_4bit=True,
    \n bnb_4bit_quant_type="nf4",
    \n bnb_4bit_compute_dtype=torch.bfloat16,
    \n bnb_4bit_use_double_quant=True # \u5f00\u542f\u53cc\u91cf\u5316\u538b\u7f29\u91cf\u5316\u5e38\u91cf\u5f00\u9500
    \n )<\/p>\n

    model = AutoModelForCausalLM.from_pretrained(
    \n model_name,
    \n quantization_config=bnb_config,
    \n device_map="auto"
    \n )<\/p>\n

    # \u5f00\u542f\u68c0\u67e5\u70b9\u4ee5\u6781\u81f4\u538b\u5e73\u524d\u5411\u6fc0\u6d3b\u5cf0\u503c
    \n model = prepare_model_for_kbit_training(
    \n model,
    \n use_gradient_checkpointing=True
    \n )<\/p>\n

    peft_config = LoraConfig(
    \n r=16,
    \n lora_alpha=16,
    \n target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    \n lora_dropout=0.05,
    \n bias="none",
    \n task_type="CAUSAL_LM"
    \n )
    \n model = get_peft_model(model, peft_config)<\/p>\n

    # 2. \u8bad\u7ec3\u8d85\u53c2\u6570\u7cbe\u5fc3\u8c03\u4f18：\u89c4\u907f PCIe \u6362\u5165\u6362\u51fa\u98a0\u7c38
    \n training_args = TrainingArguments(
    \n output_dir=output_dir,
    \n per_device_train_batch_size=1, # \u6781\u9650\u5355\u6b65\u6279\u6b21，\u538b\u4f4e\u4e2d\u95f4\u6fc0\u6d3b
    \n gradient_accumulation_steps=16, # \u4f9d\u9760\u7d2f\u52a0\u7ef4\u6301\u5b8f\u89c2 BatchSize
    \n learning_rate=2e-4,
    \n optim="paged_adamw_8bit", # \u9501\u5b9a 8-bit \u5206\u9875\u4f18\u5316\u5668
    \n warmup_ratio=0.03,
    \n lr_scheduler_type="cosine",
    \n logging_steps=10,
    \n bf16=True,
    \n gradient_checkpointing=True,
    \n # \u7981\u7528\u975e\u5fc5\u8981\u7684\u8bc4\u4f30\u4e0e\u4fdd\u5b58\u5f00\u9500\u4ee5\u91ca\u653e\u663e\u5b58\u7f13\u51b2\u533a
    \n save_strategy="steps",
    \n save_steps=500
    \n )<\/p>\n

    return model, training_args<\/p>\n

    \u4e09\u3001\u771f\u5b9e\u5fae\u8c03\u538b\u6d4b\u5bf9\u8d26：\u8bad\u7ec3\u541e\u5410\u63d0\u5347 3.4 \u500d<\/h3>\n

    \u6211\u4eec\u5728\u5355\u5f20 NVIDIA RTX 4090（24GB \u663e\u5b58）\u4e0a\u5fae\u8c03 Llama-3-70B（QLoRA 4-bit，\u5e8f\u5217\u957f\u5ea6 4096）\u3002\u5bf9\u6bd4\u9ed8\u8ba4\u7c97\u653e\u914d\u7f6e\u4e0e\u7ecf\u8fc7\u5206\u9875\u6cbb\u7406\u540e\u7684\u541e\u5410\u6570\u636e：<\/p>\n

    | \u8bad\u7ec3\u8c03\u4f18\u914d\u7f6e | \u5355\u6b65\u8017\u65f6 (Step Time) | GPU \u8ba1\u7b97\u6d3b\u8dc3\u5ea6 (SM Active) | PCIe \u5e26\u5bbd\u5360\u6ee1\u7387 | \u9884\u8ba1 10,000 \u6b65\u603b\u8017\u65f6 |
    \n| :— | :— | :— | :— | :— |
    \n| **\u9ed8\u8ba4\u7c97\u653e\u5206\u9875 (BS=4, 32-bit Paged)** | 14.8 \u79d2 | 21.5% (\u4e25\u91cd\u7b49\u5f85) | 94.2% (\u5e38\u6001\u5316\u62e5\u585e) | 41.1 \u5c0f\u65f6 (\u6548\u7387\u6781\u4f4e) |
    \n| **\u4f18\u5316\u6fc0\u6d3b\u5cf0\u503c (BS=1, 32-bit Paged)** | 7.2 \u79d2 | 54.0% | 41.0% | 20.0 \u5c0f\u65f6 |
    \n| **\u5168\u6cbb\u7406\u65b9\u6848 (BS=1, 8-bit Paged+\u7d2f\u52a0)**| **4.3 \u79d2** | **84.5% (\u9ad8\u6548\u9971\u6ee1)**| **6.5% (\u675c\u7edd\u9891\u7e41\u6362\u51fa)**| **11.9 \u5c0f\u65f6 (\u63d0\u901f 3.4\u500d)**|<\/p>\n

    \u5b9e\u6d4b\u6307\u6807\u5c55\u73b0\u4e86\u51b3\u5b9a\u6027\u7684\u6027\u80fd\u8dc3\u5347：\u5728\u5b8c\u5168\u6d88\u9664\u4e86\u4f18\u5316\u5668\u72b6\u6001\u5728 PCIe \u4e0a\u7684\u9891\u7e41\u98a0\u7c38\u540e，\u5355\u6b65\u8bad\u7ec3\u65f6\u95f4\u4ece 14.8 \u79d2\u5927\u5e45\u538b\u964d\u81f3 4.3 \u79d2，\u603b\u4f53\u8bad\u7ec3\u541e\u5410\u63d0\u5347\u4e86\u6574\u6574 3.4 \u500d；GPU \u7b97\u529b\u5229\u7528\u7387\u4ece 21.5% \u8dc3\u5347\u81f3 84.5%，\u8ba9\u539f\u672c\u8017\u65f6\u8fd1\u4e24\u5929\u7684 70B \u5355\u5361\u5fae\u8c03\u4efb\u52a1，\u5728\u4e0d\u5230\u534a\u5929\u7684\u65f6\u95f4\u5185\u9ad8\u8d28\u5b8c\u6210\u3002<\/p>\n

    \u56db\u3001\u5de5\u7a0b\u843d\u5730\u907f\u5751\u6307\u5357<\/h3>\n
  • \u8b66\u60d5 PyTorch \u663e\u5b58\u5206\u914d\u5668\u7f13\u5b58\u81a8\u80c0（CUDA Cache Bloat）：PyTorch \u9ed8\u8ba4\u4f1a\u7f13\u5b58\u5df2\u91ca\u653e\u7684\u663e\u5b58\u5757\u5907\u7528\u3002\u82e5\u8bbe\u7f6e\u4e0d\u5f53，\u8fd9\u4e9b\u7f13\u5b58\u5757\u4f1a\u6324\u538b\u7269\u7406\u663e\u5b58，\u8fc7\u65e9\u8bf1\u53d1\u9a71\u52a8\u7ea7\u5206\u9875\u3002\u5fc5\u987b\u5728\u8bad\u7ec3\u542f\u52a8\u524d\u8bbe\u7f6e\u73af\u5883\u53d8\u91cf export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True，\u8ba9\u663e\u5b58\u5206\u914d\u5668\u4ee5\u865a\u62df\u5185\u5b58\u6269\u5c55\u6bb5\u5f62\u5f0f\u5de5\u4f5c，\u4ece\u6839\u6e90\u4e0a\u51cf\u5c11\u7269\u7406\u663e\u5b58\u788e\u7247\u3002<\/li>\n
  • \u68af\u5ea6\u68c0\u67e5\u70b9\u5fc5\u987b\u914d\u5408\u8f93\u5165\u5c42\u4fdd\u5b58：\u5f00\u542f gradient_checkpointing \u65f6，\u52a1\u5fc5\u901a\u8fc7 model.enable_input_require_grads() \u4fdd\u62a4\u8f93\u5165\u5d4c\u5165\u5c42，\u5426\u5219\u5728\u524d\u5411\u53cd\u5411\u91cd\u65b0\u8ba1\u7b97\u65f6\u4f1a\u4e22\u5931\u68af\u5ea6\u4f20\u9012，\u5bfc\u81f4 LoRA \u6743\u91cd\u5b8c\u5168\u4e0d\u66f4\u65b0\u3002<\/li>\n
  • \u907f\u514d\u5728\u5faa\u73af\u4e2d\u9891\u7e41\u6253\u5370\u53c2\u6570\u5f20\u91cf：\u5728\u6bcf\u4e2a Step \u6253\u5370 tensor.item() \u6216\u4ece GPU \u62c9\u53d6\u5f20\u91cf\u5230 CPU \u505a\u65e5\u5fd7\u8bb0\u5f55，\u4f1a\u5f3a\u884c\u540c\u6b65 CUDA \u961f\u5217，\u7834\u574f\u5f02\u6b65\u8ba1\u7b97\u6d41\u6c34\u7ebf\u3002\u6240\u6709\u6307\u6807\u7edf\u8ba1\u5fc5\u987b\u5f02\u6b65\u7d2f\u52a0，\u6bcf\u9694\u6570\u5341\u6b65\u7edf\u4e00\u6c47\u603b\u4e00\u6b21\u3002<\/li>\n

    \u5206\u9875\u4f18\u5316\u5668\u8d4b\u4e88\u4e86\u6211\u4eec\u5728\u6781\u9650\u663e\u5b58\u4e0b\u6311\u6218\u8d85\u5927\u53c2\u6570\u7684\u53ef\u80fd\u6027，\u4f46\u53ea\u6709\u6d1e\u6089\u5e95\u5c42\u7684\u8bbf\u5b58\u5f00\u9500\u3001\u7528\u7cbe\u5bc6\u7684\u5de5\u7a0b\u63a7\u5236\u9a6f\u670d\u603b\u7ebf\u5ef6\u8fdf，\u624d\u80fd\u5c06\u8fd9\u79cd\u53ef\u80fd\u6027\u8f6c\u5316\u4e3a\u771f\u6b63\u9ad8\u6548\u7684\u751f\u4ea7\u529b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u5728\u6d88\u8d39\u7ea7\u5355\u5361（\u5982 RTX 4090 24GB）\u6216\u5355\u5f20 A100 \u4e0a\u5fae\u8c03 70B \u7ea7\u522b\u5927\u6a21\u578b\u65f6，**QLoRA \u5206\u9875\u4f18\u5316\u5668（Paged Optimizers）**\u88ab\u8a89\u4e3a\u7a77\u4eba\u7b97\u529b\u7684\u6551\u4e16\u4e3b\u3002\u901a\u8fc7 bitsandbytes \u63d0\u4f9b\u7684\u5206\u9875\u673a\u5236，\u5f53\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6fc0\u6d3b\u503c\u6216\u68af\u5ea6\u53d1\u751f\u77ac\u65f6\u5c16\u523a\u5bfc\u81f4\u663e\u5b58\u5373\u5c06\u6ea2\u51fa（OOM）\u65f6，\u5e95\u5c42\u9a71\u52a8\u4f1a\u81ea\u52a8\u5c06\u4f18\u5316\u5668\u72b6\u6001\u9875（Optim<\/p>\n","protected":false},"author":2,"featured_media":115054,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[74,66,2068,3677,50],"topic":[],"class_list":["post-115055","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-agent","tag-ai","tag-nlp","tag-tensorflow","tag-50"],"yoast_head":"\nPaged Optimizers \u663e\u5b58\u6362\u5165\u5ef6\u8fdf\u6cbb\u7406\uff1aQLoRA \u8bad\u7ec3\u541e\u5410\u8c03\u4f18\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/115055.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Paged Optimizers \u663e\u5b58\u6362\u5165\u5ef6\u8fdf\u6cbb\u7406\uff1aQLoRA \u8bad\u7ec3\u541e\u5410\u8c03\u4f18\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u5728\u6d88\u8d39\u7ea7\u5355\u5361（\u5982 RTX 4090 24GB）\u6216\u5355\u5f20 A100 \u4e0a\u5fae\u8c03 70B \u7ea7\u522b\u5927\u6a21\u578b\u65f6，**QLoRA \u5206\u9875\u4f18\u5316\u5668（Paged Optimizers）**\u88ab\u8a89\u4e3a\u7a77\u4eba\u7b97\u529b\u7684\u6551\u4e16\u4e3b\u3002\u901a\u8fc7 bitsandbytes \u63d0\u4f9b\u7684\u5206\u9875\u673a\u5236，\u5f53\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6fc0\u6d3b\u503c\u6216\u68af\u5ea6\u53d1\u751f\u77ac\u65f6\u5c16\u523a\u5bfc\u81f4\u663e\u5b58\u5373\u5c06\u6ea2\u51fa（OOM）\u65f6，\u5e95\u5c42\u9a71\u52a8\u4f1a\u81ea\u52a8\u5c06\u4f18\u5316\u5668\u72b6\u6001\u9875（Optim\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/115055.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-10T17:57:08+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261010175701-6aca7c6ddfa05.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"3 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/115055.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/115055.html\",\"name\":\"Paged Optimizers \u663e\u5b58\u6362\u5165\u5ef6\u8fdf\u6cbb\u7406\uff1aQLoRA \u8bad\u7ec3\u541e\u5410\u8c03\u4f18\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-10-10T17:57:08+00:00\",\"dateModified\":\"2026-10-10T17:57:08+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/115055.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/115055.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/115055.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Paged Optimizers \u663e\u5b58\u6362\u5165\u5ef6\u8fdf\u6cbb\u7406\uff1aQLoRA \u8bad\u7ec3\u541e\u5410\u8c03\u4f18\u5b9e\u6218\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Paged Optimizers \u663e\u5b58\u6362\u5165\u5ef6\u8fdf\u6cbb\u7406\uff1aQLoRA \u8bad\u7ec3\u541e\u5410\u8c03\u4f18\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/115055.html","og_locale":"zh_CN","og_type":"article","og_title":"Paged Optimizers \u663e\u5b58\u6362\u5165\u5ef6\u8fdf\u6cbb\u7406\uff1aQLoRA \u8bad\u7ec3\u541e\u5410\u8c03\u4f18\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u5728\u6d88\u8d39\u7ea7\u5355\u5361（\u5982 RTX 4090 24GB）\u6216\u5355\u5f20 A100 \u4e0a\u5fae\u8c03 70B \u7ea7\u522b\u5927\u6a21\u578b\u65f6，**QLoRA \u5206\u9875\u4f18\u5316\u5668（Paged Optimizers）**\u88ab\u8a89\u4e3a\u7a77\u4eba\u7b97\u529b\u7684\u6551\u4e16\u4e3b\u3002\u901a\u8fc7 bitsandbytes \u63d0\u4f9b\u7684\u5206\u9875\u673a\u5236，\u5f53\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6fc0\u6d3b\u503c\u6216\u68af\u5ea6\u53d1\u751f\u77ac\u65f6\u5c16\u523a\u5bfc\u81f4\u663e\u5b58\u5373\u5c06\u6ea2\u51fa（OOM）\u65f6，\u5e95\u5c42\u9a71\u52a8\u4f1a\u81ea\u52a8\u5c06\u4f18\u5316\u5668\u72b6\u6001\u9875（Optim","og_url":"https:\/\/www.wsisp.com\/helps\/115055.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-10-10T17:57:08+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20261010175701-6aca7c6ddfa05.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"3 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/115055.html","url":"https:\/\/www.wsisp.com\/helps\/115055.html","name":"Paged Optimizers \u663e\u5b58\u6362\u5165\u5ef6\u8fdf\u6cbb\u7406\uff1aQLoRA \u8bad\u7ec3\u541e\u5410\u8c03\u4f18\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-10-10T17:57:08+00:00","dateModified":"2026-10-10T17:57:08+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/115055.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/115055.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/115055.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"Paged Optimizers \u663e\u5b58\u6362\u5165\u5ef6\u8fdf\u6cbb\u7406\uff1aQLoRA \u8bad\u7ec3\u541e\u5410\u8c03\u4f18\u5b9e\u6218"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/115055","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=115055"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/115055\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/115054"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=115055"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=115055"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=115055"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=115055"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}