{"id":101358,"date":"2026-09-06T17:49:34","date_gmt":"2026-09-06T09:49:34","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/101358.html"},"modified":"2026-09-06T17:49:34","modified_gmt":"2026-09-06T09:49:34","slug":"qwen2-vl-%e5%be%ae%e8%b0%83%e5%ae%9e%e6%88%98%ef%bc%9a%e7%94%a8-llama-factory-%e8%ae%a9%e8%a7%86%e8%a7%89%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%ad%a6%e4%bc%9a%e6%97%a0%e4%ba%ba%e6%9c%ba%e8%88%aa%e6%8b%8d","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/101358.html","title":{"rendered":"Qwen2-VL \u5fae\u8c03\u5b9e\u6218\uff1a\u7528 LLaMA-Factory \u8ba9\u89c6\u89c9\u5927\u6a21\u578b\u5b66\u4f1a\u65e0\u4eba\u673a\u822a\u62cd\u76ee\u6807\u68c0\u6d4b\uff08\u9644\u907f\u5751\u6307\u5357\uff09"},"content":{"rendered":"

\u524d\u8a00：\u628a\u68c0\u6d4b\u4efb\u52a1\u4ea4\u7ed9\u89c6\u89c9\u8bed\u8a00\u6a21\u578b（VLM）\u505a，\u662f\u6700\u8fd1\u5f88\u591a\u9879\u76ee\u7684\u5c1d\u9c9c\u65b9\u5411\u2014\u2014\u4e0d\u7528\u6539\u67b6\u6784\u3001\u4e0d\u7528\u8bbe\u8ba1\u68c0\u6d4b\u5934，\u53ea\u8981\u6784\u9020\u597d\u5bf9\u8bdd\u6570\u636e，\u5927\u6a21\u578b\u5c31\u80fd"\u770b\u56fe\u8bf4\u8bdd"\u5f0f\u5730\u8f93\u51fa\u76ee\u6807\u6846\u3002\u4f46\u771f\u6b63\u4e0a\u624b\u4e4b\u540e\u4f60\u4f1a\u53d1\u73b0：\u5fae\u8c03\u80fd\u8dd1\u901a，\u4e0d\u4ee3\u8868\u63a8\u7406\u80fd\u7528\u3002\u672c\u6587\u5b8c\u6574\u8bb0\u5f55\u4e00\u6b21 Qwen2-VL-7B \u5728\u65e0\u4eba\u673a\u822a\u62cd\u56fe\u50cf\u4e0a\u7684\u5fae\u8c03\u5b9e\u8df5：\u4ece\u6570\u636e\u6784\u9020\u3001\u8bad\u7ec3\u53c2\u6570\u9009\u62e9，\u5230\u63a8\u7406\u9636\u6bb5\u7ffb\u8f66\u7684\u56db\u7c7b\u95ee\u9898\u4e0e\u5207\u56fe\u8865\u6551\u65b9\u6848，\u5168\u90e8\u8e29\u8fc7\u3001\u5168\u90e8\u7ed9\u51fa\u89e3\u6cd5\u3002\u53ef\u76f4\u63a5\u6284\u914d\u7f6e\u3002<\/p>\n

\u4e00\u3001\u9879\u76ee\u6982\u51b5<\/h3>\n\n\n\u9879\u76ee\u914d\u7f6e<\/tr>\n\n\n\n\n\n\n
\u4efb\u52a1<\/td>\n\u65e0\u4eba\u673a\u4fef\u89c6\u822a\u62cd\u57ce\u5e02\u95ee\u9898\u8bc6\u522b（7 \u7c7b\u76ee\u6807\u68c0\u6d4b）<\/td>\n<\/tr>\n
\u57fa\u7840\u6a21\u578b<\/td>\nQwen2-VL-7B-Instruct<\/td>\n<\/tr>\n
\u5fae\u8c03\u6846\u67b6<\/td>\nLLaMA-Factory（LoRA \/ SFT）<\/td>\n<\/tr>\n
\u786c\u4ef6<\/td>\nAutoDL \u5355\u5361 vGPU 32GB<\/td>\n<\/tr>\n
\u539f\u56fe\u5206\u8fa8\u7387<\/td>\n3840 \u00d7 2160<\/td>\n<\/tr>\n
\u6570\u636e\u89c4\u6a21<\/td>\n11469 \u5f20（\u6709\u76ee\u6807 9293 + \u7a7a\u6807\u7b7e 2176）<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

7 \u4e2a\u7c7b\u522b：\u5e97\u5916\u7ecf\u8425\u3001\u5360\u9053\u7ecf\u8425\u3001\u5c4b\u9876\u4e71\u5806\u7269\u6599\u3001\u4e34\u8857\u4e71\u5806\u7269\u6599\u3001\u66b4\u9732\u5783\u573e\u3001\u5de5\u5730\u626c\u5c18\u672a\u82eb\u76d6\u3001\u4ea4\u901a\u6807\u7ebf\u4e0d\u6e05\u6670\u3002\u5176\u4e2d"\u5c4b\u9876\u4e71\u5806\u7269\u6599"\u4e0e"\u4e34\u8857\u4e71\u5806\u7269\u6599"\u4f4d\u7f6e\u4e0d\u540c\u3001"\u5e97\u5916\u7ecf\u8425"\u4e0e"\u5360\u9053\u7ecf\u8425"\u6838\u5fc3\u533a\u522b\u662f\u6709\u65e0\u5b9e\u4f53\u5e97\u9762\u2014\u2014\u8fd9\u4e9b\u6613\u6df7\u6dc6\u7c7b\u7684\u8bbe\u8ba1\u4f1a\u5728 Prompt \u91cc\u663e\u5f0f\u7ed9\u51fa\u3002<\/p>\n

\u4e8c\u3001\u6570\u636e\u51c6\u5907<\/h3>\n

2.1 \u5148\u770b\u6570\u636e\u5206\u5e03，\u518d\u51b3\u5b9a\u8bad\u7ec3\u7b56\u7565<\/h4>\n\n\n\u7c7b\u522b\u7167\u7247\u6570\u6846\u6570<\/tr>\n\n\n\n\n\n\n\n\n
\u5c4b\u9876\u4e71\u5806\u7269\u6599<\/td>\n5014<\/td>\n12749<\/td>\n<\/tr>\n
\u4ea4\u901a\u6807\u7ebf\u4e0d\u6e05\u6670<\/td>\n2381<\/td>\n6551<\/td>\n<\/tr>\n
\u66b4\u9732\u5783\u573e<\/td>\n2273<\/td>\n4077<\/td>\n<\/tr>\n
\u5360\u9053\u7ecf\u8425<\/td>\n1563<\/td>\n3567<\/td>\n<\/tr>\n
\u5e97\u5916\u7ecf\u8425<\/td>\n1044<\/td>\n1508<\/td>\n<\/tr>\n
\u4e34\u8857\u4e71\u5806\u7269\u6599<\/td>\n1029<\/td>\n1412<\/td>\n<\/tr>\n
\u5de5\u5730\u626c\u5c18\u672a\u82eb\u76d6<\/td>\n306<\/td>\n506<\/td>\n<\/tr>\n
\u5408\u8ba1<\/td>\n13610<\/td>\n30370<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u4e24\u4e2a\u7ed3\u8bba\u76f4\u63a5\u51b3\u5b9a\u4e86\u540e\u9762\u7684\u5751：<\/p>\n

  • \u7c7b\u522b\u6781\u4e0d\u5747\u8861：\u6700\u5927\u7c7b\u662f\u6700\u5c0f\u7c7b\u7684 16 \u500d\u3002\u8fd9\u662f\u540e\u6765"\u6a21\u578b\u53ea\u5b66\u4f1a\u731c\u5360\u9053\u7ecf\u8425"\u7684\u76f4\u63a5\u539f\u56e0\u3002<\/li>\n
  • \u7a7a\u6807\u7b7e（\u8d1f\u6837\u672c）\u53ea\u4fdd\u7559\u4e86 20%（2176 \u5f20）\u3002\u8fd9\u662f\u540e\u6765\u8bef\u68c0\u7387\u504f\u9ad8\u7684\u76f4\u63a5\u539f\u56e0\u3002\u5efa\u8bae\u7a7a\u6807\u7b7e\u6bd4\u4f8b\u63d0\u5230 40%~50%，\u4e14\u5fc5\u987b\u662f"\u56f0\u96be\u8d1f\u6837\u672c"：\u8def\u9762\u3001\u6591\u9a6c\u7ebf\u3001\u9634\u5f71\u3001\u8def\u8fb9\u5806\u7269\u3001\u505c\u8f66\u4f4d\u3001\u5e7f\u544a\u724c\u3001\u6c34\u6e0d\u3001\u53cd\u5149\u2014\u2014\u8fd9\u4e9b\u624d\u662f\u8bef\u68c0\u7684\u771f\u6b63\u6765\u6e90\u3002<\/li>\n

    \"\"<\/p>\n

    2.2 \u6807\u6ce8\u683c\u5f0f\u8f6c\u6362<\/h4>\n

    \u539f\u59cb\u6807\u6ce8\u662f YOLO \u683c\u5f0f（\u5f52\u4e00\u5316\u7684\u4e2d\u5fc3\u70b9\u5750\u6807 + \u5bbd\u9ad8），\u800c\u00a0Qwen2-VL \u539f\u751f\u652f\u6301\u7684\u662f\u7edd\u5bf9\u50cf\u7d20\u5750\u6807 [x1, y1, x2, y2]，\u8f6c\u6362\u65f6\u6ce8\u610f\u4e24\u70b9：<\/p>\n

      \n
    • YOLO \u7684 (cx, cy, w, h) \u90fd\u662f 0~1 \u5f52\u4e00\u5316\u503c，\u8981\u5148\u4e58\u539f\u56fe\u5bbd\u9ad8\u518d\u8f6c\u5de6\u4e0a\/\u53f3\u4e0b\u89d2\u70b9；<\/li>\n
    • \u7a7a\u6807\u7b7e\u6837\u672c\u4fdd\u7559 has_issue=false \u7684\u5bf9\u8bdd（\u8f93\u51fa\u7a7a\u5217\u8868），\u8ba9\u6a21\u578b\u5b66\u4f1a"\u6ca1\u6709\u95ee\u9898\u5c31\u8f93\u51fa\u7a7a"\u3002<\/li>\n<\/ul>\n

      2.3 \u6570\u636e\u683c\u5f0f：ShareGPT + \u56fe\u50cf\u8def\u5f84<\/h4>\n

      LLaMA-Factory \u7684 VLM \u8bad\u7ec3\u6570\u636e\u7528 ShareGPT \u683c\u5f0f，images \u5b57\u6bb5\u653e\u56fe\u7247\u7684\u7edd\u5bf9\u8def\u5f84：<\/p>\n

      {
      \n "conversations": [
      \n {
      \n "from": "human",
      \n "value": "<image>\u8bf7\u5206\u6790\u8fd9\u5f20\u65e0\u4eba\u673a\u822a\u62cd\u56fe\u50cf\u2026\u2026"
      \n },
      \n {
      \n "from": "gpt",
      \n "value": "{\\\\"has_issue\\\\": true, \\\\"issue_count\\\\": 2, \\\\"issues\\\\": [{\\\\"type\\\\": \\\\"\u5360\u9053\u7ecf\u8425\\\\", \\\\"bbox\\\\": [1213, 856, 1892, 1290]}, {\\\\"type\\\\": \\\\"\u66b4\u9732\u5783\u573e\\\\", \\\\"bbox\\\\": [2431, 1520, 2688, 1745]}]}"
      \n }
      \n ],
      \n "images": ["\/root\/autodl-tmp\/VLM_Fine_tuning\/LLaMA-Factory\/data\/images\/xxx.jpg"]
      \n}<\/p>\n

      ！！！\u518d\u5728 dataset_info.json \u91cc\u6ce8\u518c urban_train \/ urban_val \u5373\u53ef\u3002<\/span><\/p>\n

      2.4 Prompt \u8bbe\u8ba1（\u8fd9\u6b65\u51b3\u5b9a\u8f93\u51fa\u683c\u5f0f\u7684\u7a33\u5b9a\u6027）<\/h4>\n

      System Prompt\u00a0\u8d1f\u8d23\u5b9a\u4e49\u4efb\u52a1\u3001\u7c7b\u522b\u8fb9\u754c\u548c\u6613\u6df7\u6dc6\u533a\u5206，User Prompt\u00a0\u8d1f\u8d23\u5f3a\u5236 JSON \u8f93\u51fa\u7ed3\u6784：（\u6211\u8fd9\u91cc\u662ffew-shot，\u53ef\u4ee5\u8ba9vlm\u505a\u53c2\u8003）<\/p>\n

      \u4f60\u662f\u4e00\u4e2a\u4e13\u4e1a\u7684\u57ce\u5e02\u7ba1\u7406\u95ee\u9898\u8bc6\u522b\u52a9\u624b，\u7528\u4e8e\u5206\u6790\u65e0\u4eba\u673a\u4fef\u89c6\u822a\u62cd\u56fe\u50cf\u3002
      \n\u56fe\u50cf\u6765\u6e90\u4e3a\u56fa\u5b9a\u9ad8\u5ea6\u7684\u65e0\u4eba\u673a\u4fef\u62cd，\u76ee\u6807\u4ee5\u4fef\u89c6\u89d2\u5ea6\u5448\u73b0，\u53ef\u80fd\u8f83\u5c0f，
      \n\u540c\u4e00\u56fe\u50cf\u4e2d\u53ef\u80fd\u5b58\u5728\u591a\u4e2a\u540c\u7c7b\u6216\u4e0d\u540c\u7c7b\u95ee\u9898，\u4e5f\u53ef\u80fd\u4e0d\u5b58\u5728\u4efb\u4f55\u95ee\u9898\u3002<\/p>\n

      \u4f60\u9700\u8981\u8bc6\u522b\u4ee5\u4e0b\u4e03\u7c7b\u57ce\u5e02\u7ba1\u7406\u95ee\u9898：
      \n1. \u5e97\u5916\u7ecf\u8425：\u6709\u5b9e\u4f53\u5e97\u9762，\u7ecf\u8425\u7269\u54c1\u4ece\u5e97\u94fa\u95e8\u53e3\u5411\u5916\u5ef6\u4f38\u6446\u653e\u2026\u2026
      \n2. \u5360\u9053\u7ecf\u8425：\u5728\u9a6c\u8def\u3001\u4eba\u884c\u9053\u6216\u5c0f\u9053\u4e0a\u6446\u653e\u7ecf\u8425，\u65e0\u5bf9\u5e94\u5b9e\u4f53\u5e97\u9762\u2026\u2026
      \n3. \u5c4b\u9876\u4e71\u5806\u7269\u6599：\u5efa\u7b51\u5c4b\u9876\u5b58\u5728\u6742\u4e71\u5806\u653e\u7684\u7269\u54c1\u2026\u2026
      \n4. \u4e34\u8857\u4e71\u5806\u7269\u6599：\u4e34\u8857\u533a\u57df\u6742\u4e71\u5806\u653e\u5efa\u7b51\u6750\u6599\u3001\u6742\u7269\u6216\u8d27\u7269\u2026\u2026
      \n5. \u66b4\u9732\u5783\u573e：\u751f\u6d3b\u5783\u573e\u3001\u5e9f\u5f03\u7269\u7b49\u5783\u573e\u7c7b\u7269\u54c1\u672a\u7ecf\u8986\u76d6\u6216\u5bb9\u7eb3\u76f4\u63a5\u66b4\u9732\u5728\u5916\u2026\u2026
      \n6. \u5de5\u5730\u626c\u5c18\u672a\u82eb\u76d6：\u5efa\u7b51\u5de5\u5730\u7684\u7269\u6599\u3001\u571f\u65b9\u6216\u6c99\u77f3\u672a\u8fdb\u884c\u8986\u76d6\u82eb\u76d6\u2026\u2026
      \n7. \u4ea4\u901a\u6807\u7ebf\u4e0d\u6e05\u6670：\u9053\u8def\u4ea4\u901a\u6807\u7ebf\u5b58\u5728\u78e8\u635f\u3001\u892a\u8272\u6216\u6a21\u7cca\u4e0d\u6e05\u2026\u2026<\/p>\n

      \u6ce8\u610f\u4e8b\u9879：
      \n– \u5e97\u5916\u7ecf\u8425\u548c\u5360\u9053\u7ecf\u8425\u7684\u6838\u5fc3\u533a\u522b\u5728\u4e8e\u6709\u65e0\u5b9e\u4f53\u5e97\u9762
      \n– \u4e34\u8857\u4e71\u5806\u7269\u6599\u548c\u66b4\u9732\u5783\u573e\u7684\u6838\u5fc3\u533a\u522b\u5728\u4e8e\u5806\u653e\u7269\u662f\u5426\u4e3a\u5783\u573e\u7c7b\u7269\u54c1
      \n– \u5c4b\u9876\u4e71\u5806\u7269\u6599\u548c\u4e34\u8857\u4e71\u5806\u7269\u6599\u7684\u533a\u522b\u5728\u4e8e\u5806\u653e\u4f4d\u7f6e\u662f\u5426\u5728\u5efa\u7b51\u5c4b\u9876
      \n\u8bf7\u4e25\u683c\u6309\u7167\u6307\u5b9aJSON\u683c\u5f0f\u8f93\u51fa\u7ed3\u679c，\u4e0d\u8981\u8f93\u51fa\u4efb\u4f55\u989d\u5916\u5185\u5bb9\u3002<\/p>\n

      \u8bf7\u5206\u6790\u8fd9\u5f20\u65e0\u4eba\u673a\u822a\u62cd\u56fe\u50cf，\u8bc6\u522b\u5176\u4e2d\u5b58\u5728\u7684\u57ce\u5e02\u7ba1\u7406\u95ee\u9898，
      \n\u5e76\u4e25\u683c\u6309\u7167\u4ee5\u4e0bJSON\u683c\u5f0f\u8f93\u51fa：
      \n{
      \n "has_issue": true\u6216false,
      \n "issue_count": \u95ee\u9898\u6570\u91cf,
      \n "issues": [
      \n { "type": "\u95ee\u9898\u7c7b\u578b\u4e2d\u6587\u540d\u79f0", "bbox": [x1, y1, x2, y2] }
      \n ]
      \n}
      \n\u5176\u4e2dbbox\u4e3a\u76ee\u6807\u5728\u56fe\u50cf\u4e2d\u7684\u7edd\u5bf9\u50cf\u7d20\u5750\u6807，[x1,y1]\u4e3a\u5de6\u4e0a\u89d2，[x2,y2]\u4e3a\u53f3\u4e0b\u89d2\u3002
      \n\u5982\u679c\u56fe\u50cf\u4e2d\u4e0d\u5b58\u5728\u4efb\u4f55\u95ee\u9898，has_issue\u8f93\u51fafalse，issue_count\u8f93\u51fa0，issues\u8f93\u51fa\u7a7a\u5217\u8868\u3002<\/p>\n

      \u51e0\u4e2a\u5b9e\u6d4b\u6709\u6548\u7684\u7ec6\u8282：<\/p>\n

        \n
      • \u6613\u6df7\u6dc6\u7c7b\u7684\u5224\u522b\u6807\u51c6\u8981\u5199\u8fdb Prompt（"\u6838\u5fc3\u533a\u522b\u5728\u4e8e\u2026\u2026"），\u6bd4\u8ba9\u6a21\u578b\u81ea\u5df1\u609f\u6709\u6548\u5f97\u591a；<\/li>\n
      • bbox \u660e\u786e\u58f0\u660e"\u7edd\u5bf9\u50cf\u7d20\u5750\u6807"，\u5e76\u5728 User Prompt \u91cc\u5b9a\u4e49\u5de6\u4e0a\/\u53f3\u4e0b\u89d2，\u51cf\u5c11\u6a21\u578b\u81ea\u7531\u53d1\u6325；<\/li>\n
      • "\u4e0d\u8981\u8f93\u51fa\u4efb\u4f55\u989d\u5916\u5185\u5bb9"\u8fd9\u53e5\u8bdd\u80fd\u663e\u8457\u964d\u4f4e JSON \u524d\u540e\u6df7\u5165\u5e9f\u8bdd\u7684\u6982\u7387\u3002<\/li>\n<\/ul>\n

        \u4e09\u3001\u8bad\u7ec3\u914d\u7f6e<\/h3>\n

        3.1 \u5173\u952e\u8bad\u7ec3\u53c2\u6570（\u8c03\u4e86\u6570\u5341\u904d\u6548\u679c\u6700\u4f73\u7248）<\/h4>\n

        ### model
        \nmodel_name_or_path: \/path\/to\/Qwen2-VL-7B-Instruct
        \ntemplate: qwen2_vl<\/p>\n

        ### method
        \nstage: sft
        \nfinetuning_type: lora
        \nlora_target: all-linear
        \nlora_rank: 64
        \nlora_alpha: 128
        \nlora_dropout: 0.05<\/p>\n

        ### \u5173\u952e：\u89e3\u51bb\u89c6\u89c9\u4fa7
        \nfreeze_vision_tower: false # \u89e3\u51bb\u89c6\u89c9\u7f16\u7801\u5668
        \nfreeze_multi_modal_projector: false # \u89e3\u51bb\u6295\u5f71\u5c42
        \nfreeze_language_model: false<\/p>\n

        ### dataset
        \ndataset: urban_train
        \neval_dataset: urban_val
        \ncutoff_len: 8192
        \nimage_max_pixels: 1003520 # \u7ea6 1120\u00d7896
        \nimage_min_pixels: 1024<\/p>\n

        ### train
        \nper_device_train_batch_size: 1
        \ngradient_accumulation_steps: 8
        \nlearning_rate: 1.0e-4
        \nlr_scheduler_type: cosine
        \nwarmup_ratio: 0.05
        \nnum_train_epochs: 3.0
        \nbf16: true
        \ngradient_checkpointing: true
        \nflash_attn: fa2
        \neval_steps: 500
        \nsave_steps: 100<\/p>\n

        3.2 \u8fd9\u4e9b\u53c2\u6570\u4e3a\u4ec0\u4e48\u8fd9\u4e48\u8bbe（\u6bcf\u4e00\u884c\u80cc\u540e\u90fd\u662f\u4e00\u6b21\u7ffb\u8f66）<\/h4>\n

        freeze_vision_tower: false\u2014\u2014\u6700\u91cd\u8981\u7684\u4e00\u884c\u3002\u00a0\u7b2c\u4e00\u7248\u8bad\u7ec3\u628a\u89c6\u89c9\u5854\u51bb\u7ed3\u4e86，\u7ed3\u679c\u6a21\u578b\u7684"\u773c\u775b"\u5b8c\u5168\u6ca1\u53c2\u4e0e\u5b66\u4e60，\u5750\u6807\u5b9a\u4f4d\u4e00\u584c\u7cca\u6d82\u3002\u89e3\u51bb\u89c6\u89c9\u7f16\u7801\u5668\u548c\u6295\u5f71\u5c42，\u8ba9\u89c6\u89c9\u7279\u5f81\u771f\u6b63\u5411\u5b9a\u4f4d\u4efb\u52a1\u5bf9\u9f50，\u662f\u5750\u6807\u7cbe\u5ea6\u63d0\u5347\u6700\u660e\u663e\u7684\u5355\u70b9\u6539\u52a8\u3002<\/p>\n

        lora_rank: 64（\u521d\u7248 16）\u3002\u00a0rank=16 \u5bf9"\u770b\u56fe\u63cf\u8ff0"\u591f\u7528，\u5bf9"\u8f93\u51fa\u50cf\u7d20\u7ea7\u5750\u6807"\u5bb9\u91cf\u660e\u663e\u4e0d\u8db3，\u6846\u7684\u5927\u5c0f\u548c\u4f4d\u7f6e\u90fd\u5f88\u968f\u610f\u3002\u5b9a\u4f4d\u4efb\u52a1\u5efa\u8bae\u4ece 32 \u8d77\u6b65，\u6761\u4ef6\u5141\u8bb8\u76f4\u63a5 64\u3002alpha \u4fdd\u6301 2 \u500d rank \u5173\u7cfb\u3002<\/p>\n

        image_max_pixels: 1003520（\u7ea6 1120\u00d7896）\u3002\u00a0\u8fd9\u662f 32GB \u663e\u5b58\u4e0b\u80fd\u7a33\u5b9a\u8dd1\u7684\u5206\u8fa8\u7387\u4e0a\u9650\u3002\u6ce8\u610f\u4e00\u4e2a\u5173\u952e\u4e8b\u5b9e：\u539f\u56fe 3840\u00d72160，\u9001\u8fdb\u6a21\u578b\u4f1a\u88ab\u4e0b\u91c7\u6837\u5230\u8fd9\u4e2a\u5c3a\u5ea6\u2014\u2014\u539f\u56fe\u91cc 10\u00d710 \u50cf\u7d20\u7684\u5c0f\u76ee\u6807，\u4e0b\u91c7\u6837\u540e\u53ea\u5269 1~2 \u4e2a patch\u3002\u8fd9\u4e2a\u77db\u76fe\u5c31\u662f\u540e\u9762\u6240\u6709\u63a8\u7406\u95ee\u9898\u7684\u6839\u6e90，\u8bad\u7ec3\u53c2\u6570\u89e3\u51b3\u4e0d\u4e86，\u53ea\u80fd\u9760\u63a8\u7406\u65b9\u6848\u8865\u6551（\u89c1\u7b2c\u4e94\u8282）\u3002<\/p>\n

        batch_size 1 \u00d7 \u68af\u5ea6\u7d2f\u79ef 8\u3002\u00a032GB \u663e\u5b58\u4e0b 7B \u6a21\u578b + \u56fe\u50cf\u8f93\u5165，\u5b9e\u9645 batch \u53ea\u80fd\u5f00 1，\u7528\u68af\u5ea6\u7d2f\u79ef\u628a\u6709\u6548 batch \u6491\u5230 8\u3002\u914d\u5408 gradient_checkpointing \u548c flash_attn: fa2，\u663e\u5b58\u624d\u538b\u5f97\u4f4f\u3002<\/p>\n

        learning_rate: 1e-4\u3002\u00a0\u9996\u8f6e\u4ece\u96f6\u8bad LoRA \u7528 1e-4 \u6ca1\u95ee\u9898；\u4f46\u5982\u679c\u662f\u4ece\u5df2\u6709 checkpoint \u7eed\u8bad，\u8fd9\u4e2a\u5b66\u4e60\u7387\u504f\u5371\u9669，\u5efa\u8bae\u964d\u5230 1e-5 ~ 3e-5\u3002<\/p>\n

        3.3 \u8bad\u7ec3\u7ed3\u679c<\/h4>\n
          \n
        • \u5355\u4e2a epoch \u7ea6 7 \u5c0f\u65f6 20 \u5206\u949f（10321 \u5f20\u8bad\u7ec3\u56fe）；<\/li>\n
        • epoch 1 \u7ed3\u675f train_loss = 0.4676；<\/li>\n
        • eval_loss \u4ece 0.4302 \u4e00\u8def\u964d\u5230 0.3771，\u6536\u655b\u5e73\u7a33；<\/li>\n
        • \u8bad\u7ec3 3 epoch（\u4e2d\u9014\u670d\u52a1\u5668\u5173\u673a\u8fc7\u4e00\u6b21，\u4ece checkpoint \u7684 LoRA \u6743\u91cd\u7eed\u8dd1，\u6b65\u6570\u4ece\u5934\u8ba1）\u3002<\/li>\n<\/ul>\n

          \"\"<\/p>\n

          \u56db\u3001\u6574\u56fe\u63a8\u7406：\u56db\u7c7b\u95ee\u9898，\u4e00\u6b21\u7ffb\u8f66\u73b0\u573a<\/h3>\n

          \u8bad\u7ec3\u6307\u6807\u5f88\u5065\u5eb7，\u4f46\u62ff\u6574\u5f20 3840\u00d72160 \u539f\u56fe\u76f4\u63a5\u63a8\u7406，\u6548\u679c\u79bb\u53ef\u7528\u8fd8\u5dee\u5f97\u8fdc：<\/p>\n

        • \u6846\u4e0d\u51c6：\u6846\u6574\u9f50\u5730\u5806\u5728\u4e00\u8d77\u3001\u5927\u5c0f\u4e00\u81f4，\u4f46\u5c31\u662f\u4e0d\u8d34\u76ee\u6807\u2014\u2014\u6a21\u578b\u5728"\u731c"\u4e00\u4e2a\u770b\u8d77\u6765\u5408\u7406\u7684\u5750\u6807，\u800c\u4e0d\u662f\u5728\u505a\u51e0\u4f55\u56de\u5f52；<\/li>\n
        • \u6846\u504f\u5927：\u666e\u904d\u6bd4\u771f\u5b9e\u76ee\u6807\u5927 3~4 \u500d；<\/li>\n
        • \u8bef\u68c0：\u628a\u9a6c\u8def\u7eb9\u7406\u3001\u9634\u5f71\u68c0\u6210"\u66b4\u9732\u5783\u573e""\u5360\u9053\u7ecf\u8425"；<\/li>\n
        • \u6f0f\u68c0：\u5c0f\u76ee\u6807\u5bc6\u96c6\u7684\u56fe，\u68c0\u51fa\u6570\u91cf\u660e\u663e\u4e0d\u8db3\u3002<\/li>\n

          \u6839\u56e0\u5206\u6790（\u8fd9\u90e8\u5206\u662f\u5168\u6587\u6700\u503c\u94b1\u7684\u8ba4\u77e5）<\/h4>\n

          VLM \u4e0d\u662f\u68c0\u6d4b\u6a21\u578b\u3002\u00a0\u4f20\u7edf\u68c0\u6d4b\u5668（YOLO\u3001DINO）\u6709 FPN \u591a\u5c3a\u5ea6\u7279\u5f81\u91d1\u5b57\u5854\u548c anchor \u673a\u5236\u4e13\u6cbb\u5c0f\u76ee\u6807；Qwen2-VL \u7684\u89c6\u89c9\u7f16\u7801\u5668\u628a\u4efb\u610f\u5c3a\u5bf8\u8f93\u5165\u6620\u5c04\u5230\u56fa\u5b9a\u6570\u91cf\u7684 patch token，\u5c0f\u76ee\u6807\u5728\u4e0b\u91c7\u6837\u540e\u7279\u5f81\u51e0\u4e4e\u88ab\u6df9\u6ca1\u3002\u5b83\u8f93\u51fa bbox \u7684\u65b9\u5f0f\u662f\u8ba9\u8bed\u8a00\u6a21\u578b\u628a\u5750\u6807"\u8bf4"\u51fa\u6765\u2014\u2014\u672c\u8d28\u662f\u5e8f\u5217\u9884\u6d4b，\u4e0d\u662f\u51e0\u4f55\u56de\u5f52\u3002\u6846\u6574\u9f50\u6392\u5217\u3001\u5927\u5c0f\u4e00\u81f4，\u6b63\u662f\u8bed\u8a00\u6a21\u578b\u5728\u6309"\u7edf\u8ba1\u89c4\u5f8b"\u8f93\u51fa\u5750\u6807\u683c\u5f0f\u7684\u5178\u578b\u75c7\u72b6\u3002<\/p>\n

          \u5269\u4e0b\u7684\u95ee\u9898\u6765\u81ea\u6570\u636e：<\/p>\n

            \n
          • \u8bef\u68c0：\u7a7a\u6807\u7b7e\u53ea\u5360 20%，\u4e14\u7f3a"\u8def\u9762\/\u9634\u5f71\/\u6591\u9a6c\u7ebf"\u8fd9\u7c7b\u56f0\u96be\u8d1f\u6837\u672c；<\/li>\n
          • \u7c7b\u522b\u504f\u79d1：\u7b2c\u4e00\u8f6e\u8bad\u7ec3 1 \u4e2a epoch \u65f6\u6a21\u578b\u53ea\u8f93\u51fa"\u5360\u9053\u7ecf\u8425"\u2014\u2014\u591a\u6570\u7c7b\u6837\u672c\u591a，\u6a21\u578b\u5b66\u5230\u4e86"\u731c\u591a\u6570\u7c7b\u6c38\u8fdc\u4e0d\u4f1a\u592a\u5dee"\u7684\u6377\u5f84，\u5178\u578b\u7684\u7c7b\u522b\u4e0d\u5e73\u8861 + \u6b20\u8bad\u7ec3\u7ec4\u5408\u3002<\/li>\n<\/ul>\n

            \"\"
            \n\u793a\u4f8b：\u91cd\u53e0\u5f88\u591a<\/p>\n

            \u4e94\u3001\u5207\u56fe\u63a8\u7406：\u4e0d\u6539\u4e00\u884c\u8bad\u7ec3\u4ee3\u7801，\u6548\u679c\u5927\u5e45\u63d0\u5347<\/h3>\n

            \u65e2\u7136\u5c0f\u76ee\u6807\u7684\u77db\u76fe\u662f"\u4e0b\u91c7\u6837\u540e\u7279\u5f81\u88ab\u6df9\u6ca1"，\u90a3\u5c31\u5728\u63a8\u7406\u4fa7\u628a\u76ee\u6807\u5728\u8f93\u5165\u56fe\u4e2d\u7684\u76f8\u5bf9\u5c3a\u5bf8\u653e\u5927\u2014\u2014\u628a\u539f\u56fe\u5207\u5757，\u9010\u5757\u63a8\u7406，\u5750\u6807\u8fd8\u539f\u540e\u5408\u5e76\u53bb\u91cd\u3002<\/p>\n

            5.1 \u65b9\u6848\u8bbe\u8ba1<\/h4>\n
              \n
            • \u539f\u56fe 3840\u00d72160 \u5207\u6210\u00a02\u00d73 = 6 \u5757，\u6bcf\u5757\u7ea6 1280\u00d7960，\u76ee\u6807\u76f8\u5bf9\u5c3a\u5bf8\u653e\u5927\u7ea6 2.4 \u500d；<\/li>\n
            • \u5757\u95f4\u4fdd\u7559\u00a015% overlap，\u9632\u6b62\u76ee\u6807\u6b63\u597d\u538b\u5728\u5207\u5272\u8fb9\u754c\u4e0a\u88ab\u622a\u65ad；<\/li>\n
            • \u5c0f\u76ee\u6807\u7279\u522b\u5bc6\u96c6\u7684\u56fe\u8fdb\u4e00\u6b65\u5207\u6210\u00a015 \u5757；<\/li>\n
            • \u6bcf\u5757\u72ec\u7acb\u63a8\u7406\u540e，\u628a\u5750\u6807\u5f52\u4e00\u5316 \u2192 \u6620\u5c04\u56de\u5b50\u56fe\u50cf\u7d20\u5c3a\u5ea6 \u2192 \u52a0\u5b50\u56fe\u504f\u79fb\u91cf，\u8fd8\u539f\u4e3a\u539f\u56fe\u5750\u6807；<\/li>\n
            • \u6700\u540e\u00a0NMS \u53bb\u91cd，\u53bb\u6389 overlap \u533a\u57df\u4ea7\u751f\u7684\u91cd\u590d\u6846\u3002<\/li>\n<\/ul>\n

              5.2 \u6838\u5fc3\u4ee3\u7801<\/h4>\n

              import cv2
              \nimport numpy as np<\/p>\n

              def slice_inference(image, predict_fn, rows=2, cols=3, overlap=0.15):
              \n """\u5207\u56fe\u63a8\u7406：\u9010\u5757\u63a8\u7406 + \u5750\u6807\u8fd8\u539f + NMS \u5408\u5e76"""
              \n H, W = image.shape[:2]
              \n # \u5e26\u91cd\u53e0\u7684\u5207\u5757：\u6bcf\u5757\u7684\u5b9e\u9645\u8d77\u6b62\u8303\u56f4
              \n ys = np.linspace(0, H, rows + 1).astype(int)
              \n xs = np.linspace(0, W, cols + 1).astype(int)
              \n pad_y, pad_x = int(H * overlap), int(W * overlap)<\/p>\n

              all_boxes, all_scores, all_labels = [], [], []
              \n for i in range(rows):
              \n for j in range(cols):
              \n y0, y1 = max(0, ys[i] – pad_y), min(H, ys[i + 1] + pad_y)
              \n x0, x1 = max(0, xs[j] – pad_x), min(W, xs[j + 1] + pad_x)
              \n tile = image[y0:y1, x0:x1]
              \n tw, th = tile.shape[1], tile.shape[0]<\/p>\n

              for box, score, label in predict_fn(tile):
              \n x1_, y1_, x2_, y2_ = box
              \n # \u6a21\u578b\u5076\u5c14\u8f93\u51fa 0~1 \u76f8\u5bf9\u5750\u6807，\u7edf\u4e00\u8f6c\u6210\u5b50\u56fe\u50cf\u7d20\u5750\u6807
              \n if max(x1_, y1_, x2_, y2_) <= 1.5:
              \n x1_, y1_, x2_, y2_ = x1_*tw, y1_*th, x2_*tw, y2_*th
              \n # \u52a0\u4e0a\u5b50\u56fe\u5de6\u4e0a\u89d2\u504f\u79fb，\u8fd8\u539f\u5230\u539f\u56fe\u5750\u6807\u7cfb
              \n all_boxes.append([x1_+x0, y1_+y0, x2_+x0, y2_+y0])
              \n all_scores.append(score)
              \n all_labels.append(label)<\/p>\n

              keep = nms(all_boxes, all_scores, iou_thr=0.5) # \u53bb\u6389\u91cd\u53e0\u533a\u91cd\u590d\u6846
              \n return [(all_labels[k], all_boxes[k]) for k in keep]<\/p>\n

              \u5750\u6807\u8fd8\u539f\u7684\u5b8c\u6574\u94fe\u8def：\u6a21\u578b\u8f93\u51fa（\u76f8\u5bf9\u5b50\u56fe）\u2192 \u5f52\u4e00\u5316\u5224\u65ad \u2192 \u4e58\u5b50\u56fe\u5b9e\u9645\u5c3a\u5bf8 \u2192 \u52a0\u504f\u79fb\u91cf \u2192 \u539f\u56fe\u5750\u6807 \u2192 NMS\u3002\u5b9e\u6d4b\u4e2d\u6a21\u578b\u8f93\u51fa\u5750\u6807\u5236\u5f0f\u5076\u5c14\u6f02\u79fb（\u6709\u65f6\u7edd\u5bf9\u50cf\u7d20\u3001\u6709\u65f6 0~1 \u76f8\u5bf9\u5750\u6807），\u6240\u4ee5\u4ee3\u7801\u91cc\u4fdd\u7559\u4e86\u4e00\u6b21 <=1.5 \u7684\u5224\u65ad\u505a\u81ea\u9002\u5e94\u3002<\/p>\n

              5.3 \u6548\u679c<\/h4>\n

              \u5c0f\u76ee\u6807\u7684\u5b9a\u4f4d\u4e0e\u8bc6\u522b\u7cbe\u5ea6\u660e\u663e\u6539\u5584，\u5bc6\u96c6\u5c0f\u76ee\u6807\u573a\u666f\u7684\u6f0f\u68c0\u4e5f\u5927\u5e45\u7f13\u89e3，\u6574\u4f53\u8fbe\u5230\u53ef\u6f14\u793a\u6c34\u5e73\u3002\u4ee3\u4ef7\u662f\u63a8\u7406\u6b21\u6570\u53d8\u6210 6~15 \u500d，\u9700\u8981\u6309\u4e1a\u52a1\u5ef6\u8fdf\u505a\u6743\u8861\u3002<\/p>\n

              \"\"\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u00a0 \u00a0\u00a0\u00a0\"\"<\/p>\n

              \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u5bc6\u96c6\u5c0f\u76ee\u6807\u573a\u666f，\u5207 6 \u5757 vs \u5207 15 \u5757\u5bf9\u6bd4<\/p>\n

              \u516d\u3001\u907f\u5751\u6e05\u5355（\u90fd\u662f\u771f\u91d1\u767d\u94f6\u6362\u6765\u7684）<\/h3>\n
            • \u5750\u6807\u7cfb\u8981\u5168\u94fe\u8def\u5bf9\u9f50\u3002\u8bad\u7ec3\u6570\u636e\u662f\u539f\u56fe\u7edd\u5bf9\u50cf\u7d20\u5750\u6807，\u6a21\u578b\u770b\u5230\u7684\u662f\u7f29\u653e\u540e\u7684\u56fe；\u63a8\u7406\u65f6\u6a21\u578b\u8f93\u51fa\u6709\u65f6\u662f 0~1 \u76f8\u5bf9\u5750\u6807\u2014\u2014\u6bcf\u4e2a\u73af\u8282\u90fd\u663e\u5f0f\u786e\u8ba4\u5750\u6807\u5236\u5f0f，\u80fd\u7701\u6389\u4e00\u6574\u5929\u7684 debug\u3002<\/li>\n
            • \u89c6\u89c9\u5854\u5fc5\u987b\u89e3\u51bb\u3002freeze_vision_tower: false，\u4e0d\u7136\u6a21\u578b\u7684"\u773c\u775b"\u4e0d\u53c2\u4e0e\u5b66\u4e60，\u5b9a\u4f4d\u5168\u9760\u8499\u3002<\/li>\n
            • LoRA rank \u522b\u4f4e\u4e8e 32\u3002\u5b9a\u4f4d\u4efb\u52a1\u5bf9\u5bb9\u91cf\u6bd4\u63cf\u8ff0\u4efb\u52a1\u654f\u611f\u5f97\u591a，16 \u4f1a\u660e\u663e\u4e0d\u591f\u7528\u3002<\/li>\n
            • \u7a7a\u6807\u7b7e\u4e0d\u662f\u5783\u573e\u6570\u636e，\u662f\u538b\u5236\u8bef\u68c0\u7684\u4e3b\u529b\u300220% \u4e0d\u591f，40%~50% \u66f4\u7a33，\u4e14\u4f18\u5148\u8865\u5145"\u957f\u5f97\u50cf\u76ee\u6807"\u7684\u56f0\u96be\u8d1f\u6837\u672c\u3002<\/li>\n
            • \u7c7b\u522b\u4e0d\u5e73\u8861\u8981\u5728\u6570\u636e\u5c42\u5904\u7406（\u8fc7\u91c7\u6837\u5c11\u6570\u7c7b \/ \u52a0\u6743 loss），\u5426\u5219\u6a21\u578b\u4f1a\u8d70"\u731c\u591a\u6570\u7c7b"\u7684\u6377\u5f84\u3002<\/li>\n
            • \u5bc6\u96c6\u76ee\u6807\u7684 JSON \u8f93\u51fa\u53ef\u80fd\u88ab max_new_tokens \u622a\u65ad（\u4e00\u56fe 15 \u4e2a\u76ee\u6807\u7684 JSON \u5f88\u5bb9\u6613\u8d85 1024 token），\u6f0f\u68c0\u6392\u67e5\u65f6\u8bb0\u5f97\u68c0\u67e5\u8fd9\u4e00\u9879\u2014\u2014\u867d\u7136\u591a\u6570\u65f6\u5019\u4e3b\u56e0\u8fd8\u662f\u5c0f\u76ee\u6807\u7279\u5f81\u592a\u5f31，\u6a21\u578b\u6839\u672c\u6ca1\u770b\u89c1\u3002<\/li>\n
            • OpenCV \u753b\u4e2d\u6587\u6807\u7b7e\u4f1a\u53d8\u65b9\u5757\u3002\u670d\u52a1\u5668\u88c5 fonts-wqy-zenhei，\u6216\u8005\u6539\u7528 PIL \u7ed8\u5236\u4e2d\u6587\u3002<\/li>\n
            • \u957f\u8bad\u4f1a\u9047\u4e0a\u65ad\u70b9\u7eed\u8bad\u3002\u4e91\u7aef\u5b9e\u4f8b\u5173\u673a\u4e2d\u65ad\u5f88\u5e38\u89c1，\u4ece checkpoint \u7684 LoRA \u6743\u91cd\u7eed\u8dd1\u5373\u53ef；\u7eed\u8bad\u65f6\u628a\u5b66\u4e60\u7387\u964d\u5230 1e-5 ~ 3e-5，\u9632\u6b62\u628a\u5df2\u6536\u655b\u7684\u6743\u91cd\u51b2\u98de\u3002<\/li>\n
            • \u8bad\u7ec3 loss \u6b63\u5e38 \u2260 \u63a8\u7406\u53ef\u7528\u3002SFT \u6536\u655b\u53ea\u8bf4\u660e\u6a21\u578b\u5b66\u4f1a\u4e86\u8f93\u51fa\u683c\u5f0f，VLM \u505a\u68c0\u6d4b\u7684\u5929\u82b1\u677f\u7531\u67b6\u6784\u51b3\u5b9a：\u6ca1\u6709 FPN，\u5c31\u6ca1\u6709\u5c0f\u76ee\u6807\u7cbe\u5ea6\u4e0a\u9650\u7684\u4fdd\u8bc1\u3002<\/li>\n

              \u4e03\u3001\u540e\u7eed\u4f18\u5316\u65b9\u5411<\/h3>\n

              \u6309\u6295\u5165\u4ea7\u51fa\u6392\u5e8f：<\/p>\n

            • \u63a8\u7406\u540e\u5904\u7406（\u96f6\u8bad\u7ec3\u6210\u672c）：NMS \u7684 IoU \u9608\u503c\u4ece 0.5 \u6536\u7d27\u5230 0.35；\u6309\u9762\u79ef\u8fc7\u6ee4\u660e\u663e\u5f02\u5e38\u7684\u6846（\u8d85\u8fc7\u539f\u56fe 30% \u57fa\u672c\u662f\u8bef\u68c0，\u5c0f\u4e8e 0.01% \u57fa\u672c\u662f\u566a\u58f0）\u3002<\/li>\n
            • YOLO \u4e24\u9636\u6bb5\u6d41\u6c34\u7ebf（\u6700\u63a8\u8350）：\u7528\u73b0\u6210\u7684 YOLO \u683c\u5f0f\u6807\u6ce8\u76f4\u63a5\u8bad\u4e00\u4e2a YOLOv8，\u8d1f\u8d23"\u5728\u54ea\u91cc"；Qwen2-VL \u53ea\u5bf9 YOLO \u6846\u51fa\u7684 crop \u505a\u5206\u7c7b\u786e\u8ba4，\u8d1f\u8d23"\u662f\u4ec0\u4e48"\u3002\u68c0\u6d4b\u5668\u7684 FPN \u7ed3\u6784\u5929\u7136\u9002\u914d\u5c0f\u76ee\u6807，VLM \u7684\u8bed\u4e49\u7406\u89e3\u80fd\u529b\u7528\u5728\u5200\u5203\u4e0a，\u73b0\u6709\u6807\u6ce8\u96f6\u989d\u5916\u5904\u7406\u3002<\/li>\n
            • GRPO \u5f3a\u5316\u8bad\u7ec3（\u957f\u671f）：\u5728 SFT \u57fa\u7840\u4e0a\u8bbe\u8ba1 IoU \u5956\u52b1 + \u5206\u7c7b\u5956\u52b1 + \u683c\u5f0f\u5956\u52b1，\u8ba9\u5b9a\u4f4d\u4ece"\u8bed\u8a00\u6a21\u578b\u731c\u5750\u6807"\u8fdb\u5316\u5230"\u771f\u6b63\u4f18\u5316\u51e0\u4f55\u7cbe\u5ea6"\u3002\u524d\u63d0\u662f\u5148\u505a rule-based reward \u79bb\u7ebf\u9a8c\u8bc1，\u5426\u5219 bbox \u683c\u5f0f\u8fd8\u6f02\u7740\u5c31\u4e0a RL，\u4f1a\u76f4\u63a5\u5b66\u6b6a\u3002<\/li>\n

              \u603b\u7ed3<\/h3>\n

              \u8fd9\u6b21\u5b9e\u8df5\u5b8c\u6574\u7684\u8def\u7ebf\u662f：ShareGPT \u683c\u5f0f\u6784\u9020\u5e26\u4e25\u683c JSON \u8f93\u51fa\u6a21\u677f\u7684\u68c0\u6d4b\u6570\u636e \u2192 LLaMA-Factory \u4e0a\u89e3\u51bb\u89c6\u89c9\u5854\u7684 LoRA \u5fae\u8c03（rank=64）\u2192 \u9762\u5bf9\u9ad8\u5206\u8fa8\u7387\u5c0f\u76ee\u6807，\u7528\u5207\u56fe\u63a8\u7406 + \u5750\u6807\u8fd8\u539f + NMS \u8865\u6551\u3002<\/p>\n

              \u6700\u60f3\u7559\u7ed9\u4f60\u7684\u4e00\u53e5\u8bdd：VLM \u505a\u68c0\u6d4b，\u5fae\u8c03\u89e3\u51b3\u7684\u662f"\u8f93\u51fa\u683c\u5f0f\u548c\u8bed\u4e49\u5bf9\u9f50"，\u89e3\u51b3\u4e0d\u4e86"\u5c0f\u76ee\u6807\u51e0\u4f55\u5b9a\u4f4d"\u2014\u2014\u540e\u8005\u8981\u4e48\u9760\u5207\u56fe\u63a8\u7406\u7f13\u89e3，\u8981\u4e48\u8001\u5b9e\u4e0a\u4e24\u9636\u6bb5\u6d41\u6c34\u7ebf\u3002\u8ba4\u6e05\u8fd9\u4e2a\u8fb9\u754c，\u80fd\u5e2e\u4f60\u7701\u4e0b\u5927\u91cf\u8c03\u53c2\u65f6\u95f4\u3002<\/p>\n

              \u5982\u679c\u8fd9\u7bc7\u5bf9\u4f60\u6709\u5e2e\u52a9，\u6b22\u8fce\u70b9\u8d5e\u6536\u85cf；\u8e29\u8fc7\u522b\u7684\u5751\u4e5f\u53ef\u4ee5\u5728\u8bc4\u8bba\u533a\u4ea4\u6d41\u3002\u6211\u505a\u7684\u4e0d\u4e00\u5b9a\u662f\u6700\u597d\u7684，\u6b22\u8fce\u5927\u5bb6\u8ba8\u8bba~<\/p>\n

              \n

              \u73af\u5883\u8bf4\u660e：Qwen2-VL-7B-Instruct \/ LLaMA-Factory \/ AutoDL vGPU 32GB \/ PyTorch + flash-attn<\/p>\n","protected":false},"excerpt":{"rendered":"

              \u524d\u8a00：\u628a\u68c0\u6d4b\u4efb\u52a1\u4ea4\u7ed9\u89c6\u89c9\u8bed\u8a00\u6a21\u578b（VLM）\u505a，\u662f\u6700\u8fd1\u5f88\u591a\u9879\u76ee\u7684\u5c1d\u9c9c\u65b9\u5411\u2014\u2014\u4e0d\u7528\u6539\u67b6\u6784\u3001\u4e0d\u7528\u8bbe\u8ba1\u68c0\u6d4b\u5934，\u53ea\u8981\u6784\u9020\u597d\u5bf9\u8bdd\u6570\u636e，\u5927\u6a21\u578b\u5c31\u80fd\\”\u770b\u56fe\u8bf4\u8bdd\\”\u5f0f\u5730\u8f93\u51fa\u76ee\u6807\u6846\u3002\u4f46\u771f\u6b63\u4e0a\u624b\u4e4b\u540e\u4f60\u4f1a\u53d1\u73b0：\u5fae\u8c03\u80fd\u8dd1\u901a，\u4e0d\u4ee3\u8868\u63a8\u7406\u80fd\u7528\u3002\u672c\u6587\u5b8c\u6574\u8bb0\u5f55\u4e00\u6b21 Qwen2-VL-7B \u5728\u65e0\u4eba\u673a\u822a\u62cd\u56fe\u50cf\u4e0a\u7684\u5fae\u8c03\u5b9e\u8df5：\u4ece\u6570\u636e\u6784\u9020\u3001\u8bad\u7ec3\u53c2\u6570\u9009\u62e9，\u5230\u63a8\u7406\u9636\u6bb5\u7ffb\u8f66\u7684\u56db\u7c7b\u95ee\u9898\u4e0e\u5207\u56fe<\/p>\n","protected":false},"author":2,"featured_media":101353,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[347,50,1665],"topic":[],"class_list":["post-101358","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-llama","tag-50","tag-1665"],"yoast_head":"\nQwen2-VL \u5fae\u8c03\u5b9e\u6218\uff1a\u7528 LLaMA-Factory \u8ba9\u89c6\u89c9\u5927\u6a21\u578b\u5b66\u4f1a\u65e0\u4eba\u673a\u822a\u62cd\u76ee\u6807\u68c0\u6d4b\uff08\u9644\u907f\u5751\u6307\u5357\uff09 - \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\/101358.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Qwen2-VL \u5fae\u8c03\u5b9e\u6218\uff1a\u7528 LLaMA-Factory \u8ba9\u89c6\u89c9\u5927\u6a21\u578b\u5b66\u4f1a\u65e0\u4eba\u673a\u822a\u62cd\u76ee\u6807\u68c0\u6d4b\uff08\u9644\u907f\u5751\u6307\u5357\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u524d\u8a00：\u628a\u68c0\u6d4b\u4efb\u52a1\u4ea4\u7ed9\u89c6\u89c9\u8bed\u8a00\u6a21\u578b（VLM）\u505a，\u662f\u6700\u8fd1\u5f88\u591a\u9879\u76ee\u7684\u5c1d\u9c9c\u65b9\u5411\u2014\u2014\u4e0d\u7528\u6539\u67b6\u6784\u3001\u4e0d\u7528\u8bbe\u8ba1\u68c0\u6d4b\u5934，\u53ea\u8981\u6784\u9020\u597d\u5bf9\u8bdd\u6570\u636e，\u5927\u6a21\u578b\u5c31\u80fd"\u770b\u56fe\u8bf4\u8bdd"\u5f0f\u5730\u8f93\u51fa\u76ee\u6807\u6846\u3002\u4f46\u771f\u6b63\u4e0a\u624b\u4e4b\u540e\u4f60\u4f1a\u53d1\u73b0：\u5fae\u8c03\u80fd\u8dd1\u901a，\u4e0d\u4ee3\u8868\u63a8\u7406\u80fd\u7528\u3002\u672c\u6587\u5b8c\u6574\u8bb0\u5f55\u4e00\u6b21 Qwen2-VL-7B \u5728\u65e0\u4eba\u673a\u822a\u62cd\u56fe\u50cf\u4e0a\u7684\u5fae\u8c03\u5b9e\u8df5：\u4ece\u6570\u636e\u6784\u9020\u3001\u8bad\u7ec3\u53c2\u6570\u9009\u62e9，\u5230\u63a8\u7406\u9636\u6bb5\u7ffb\u8f66\u7684\u56db\u7c7b\u95ee\u9898\u4e0e\u5207\u56fe\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/101358.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-06T09:49:34+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260906094933-6a9d372d14b55.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=\"5 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/101358.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/101358.html\",\"name\":\"Qwen2-VL \u5fae\u8c03\u5b9e\u6218\uff1a\u7528 LLaMA-Factory \u8ba9\u89c6\u89c9\u5927\u6a21\u578b\u5b66\u4f1a\u65e0\u4eba\u673a\u822a\u62cd\u76ee\u6807\u68c0\u6d4b\uff08\u9644\u907f\u5751\u6307\u5357\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-06T09:49:34+00:00\",\"dateModified\":\"2026-09-06T09:49:34+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/101358.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/101358.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/101358.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Qwen2-VL \u5fae\u8c03\u5b9e\u6218\uff1a\u7528 LLaMA-Factory \u8ba9\u89c6\u89c9\u5927\u6a21\u578b\u5b66\u4f1a\u65e0\u4eba\u673a\u822a\u62cd\u76ee\u6807\u68c0\u6d4b\uff08\u9644\u907f\u5751\u6307\u5357\uff09\"}]},{\"@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":"Qwen2-VL \u5fae\u8c03\u5b9e\u6218\uff1a\u7528 LLaMA-Factory \u8ba9\u89c6\u89c9\u5927\u6a21\u578b\u5b66\u4f1a\u65e0\u4eba\u673a\u822a\u62cd\u76ee\u6807\u68c0\u6d4b\uff08\u9644\u907f\u5751\u6307\u5357\uff09 - \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\/101358.html","og_locale":"zh_CN","og_type":"article","og_title":"Qwen2-VL \u5fae\u8c03\u5b9e\u6218\uff1a\u7528 LLaMA-Factory \u8ba9\u89c6\u89c9\u5927\u6a21\u578b\u5b66\u4f1a\u65e0\u4eba\u673a\u822a\u62cd\u76ee\u6807\u68c0\u6d4b\uff08\u9644\u907f\u5751\u6307\u5357\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u524d\u8a00：\u628a\u68c0\u6d4b\u4efb\u52a1\u4ea4\u7ed9\u89c6\u89c9\u8bed\u8a00\u6a21\u578b（VLM）\u505a，\u662f\u6700\u8fd1\u5f88\u591a\u9879\u76ee\u7684\u5c1d\u9c9c\u65b9\u5411\u2014\u2014\u4e0d\u7528\u6539\u67b6\u6784\u3001\u4e0d\u7528\u8bbe\u8ba1\u68c0\u6d4b\u5934，\u53ea\u8981\u6784\u9020\u597d\u5bf9\u8bdd\u6570\u636e，\u5927\u6a21\u578b\u5c31\u80fd\"\u770b\u56fe\u8bf4\u8bdd\"\u5f0f\u5730\u8f93\u51fa\u76ee\u6807\u6846\u3002\u4f46\u771f\u6b63\u4e0a\u624b\u4e4b\u540e\u4f60\u4f1a\u53d1\u73b0：\u5fae\u8c03\u80fd\u8dd1\u901a，\u4e0d\u4ee3\u8868\u63a8\u7406\u80fd\u7528\u3002\u672c\u6587\u5b8c\u6574\u8bb0\u5f55\u4e00\u6b21 Qwen2-VL-7B \u5728\u65e0\u4eba\u673a\u822a\u62cd\u56fe\u50cf\u4e0a\u7684\u5fae\u8c03\u5b9e\u8df5：\u4ece\u6570\u636e\u6784\u9020\u3001\u8bad\u7ec3\u53c2\u6570\u9009\u62e9，\u5230\u63a8\u7406\u9636\u6bb5\u7ffb\u8f66\u7684\u56db\u7c7b\u95ee\u9898\u4e0e\u5207\u56fe","og_url":"https:\/\/www.wsisp.com\/helps\/101358.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-06T09:49:34+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260906094933-6a9d372d14b55.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"5 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/101358.html","url":"https:\/\/www.wsisp.com\/helps\/101358.html","name":"Qwen2-VL \u5fae\u8c03\u5b9e\u6218\uff1a\u7528 LLaMA-Factory \u8ba9\u89c6\u89c9\u5927\u6a21\u578b\u5b66\u4f1a\u65e0\u4eba\u673a\u822a\u62cd\u76ee\u6807\u68c0\u6d4b\uff08\u9644\u907f\u5751\u6307\u5357\uff09 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-06T09:49:34+00:00","dateModified":"2026-09-06T09:49:34+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/101358.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/101358.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/101358.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"Qwen2-VL \u5fae\u8c03\u5b9e\u6218\uff1a\u7528 LLaMA-Factory \u8ba9\u89c6\u89c9\u5927\u6a21\u578b\u5b66\u4f1a\u65e0\u4eba\u673a\u822a\u62cd\u76ee\u6807\u68c0\u6d4b\uff08\u9644\u907f\u5751\u6307\u5357\uff09"}]},{"@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\/101358","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=101358"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/101358\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/101353"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=101358"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=101358"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=101358"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=101358"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}