{"id":111371,"date":"2026-10-01T01:09:47","date_gmt":"2026-09-30T17:09:47","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/111371.html"},"modified":"2026-10-01T01:09:47","modified_gmt":"2026-09-30T17:09:47","slug":"%e8%a7%86%e9%a2%91%e7%90%86%e8%a7%a3%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%ae%9e%e6%88%98%ef%bc%9a%e7%94%a8-qwen3-vl-%e5%81%9a%e9%ab%98%e5%85%89%e7%89%87%e6%ae%b5%e5%ae%9a%e4%bd%8d%e4%b8%8e%e9%80%90%e5%b8%a7","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/111371.html","title":{"rendered":"\u89c6\u9891\u7406\u89e3\u5927\u6a21\u578b\u5b9e\u6218\uff1a\u7528 Qwen3-VL \u505a\u9ad8\u5149\u7247\u6bb5\u5b9a\u4f4d\u4e0e\u9010\u5e27\u6784\u56fe"},"content":{"rendered":"

\u4ece\u4e00\u6bb5\u957f\u89c6\u9891\u91cc\u81ea\u52a8\u6311\u51fa\u7cbe\u5f69\u7247\u6bb5\u5e76\u7ed9\u51fa\u88c1\u526a\u6846，\u4f20\u7edf\u505a\u6cd5\u8981\u5806\u4e00\u5806\u4e13\u7528\u5c0f\u6a21\u578b\u3002\u672c\u6587\u7528 Qwen3-VL（Qwen3-VL-8B-Instruct）\u8d70\u901a\u300c\u65f6\u95f4\u6233\u5bf9\u9f50 \u2192 \u65f6\u5e8f\u5b9a\u4f4d \u2192 \u6784\u56fe\u6846\u8f93\u51fa\u300d\u7684\u5b8c\u6574\u94fe\u8def，\u9644\u53ef\u8fd0\u884c\u4ee3\u7801\u4e0e\u5b9e\u6d4b\u5f00\u9500\u3002<\/p>\n

\n

\u4e00\u3001\u5148\u8bf4\u6e05\u8981\u89e3\u51b3\u4ec0\u4e48<\/h3>\n

\u5047\u8bbe\u4f60\u6709\u4e00\u6bb5\u4e09\u5206\u949f\u7684\u7403\u8d5b\u5f55\u50cf，\u4e1a\u52a1\u4fa7\u8981\u4e00\u6761 15 \u79d2\u7684\u7ad6\u5c4f\u77ed\u7247\u3002\u8fd9\u4ef6\u4e8b\u62c6\u5f00\u6765\u662f\u4e09\u4e2a\u72ec\u7acb\u95ee\u9898：<\/p>\n

  • \u9009\u54ea\u51e0\u5e27：\u65f6\u5e8f\u5b9a\u4f4d（temporal grounding），\u8f93\u51fa\u4e00\u4e2a\u65f6\u95f4\u533a\u95f4<\/li>\n
  • \u6bcf\u5e27\u6846\u54ea\u91cc：\u6784\u56fe\u51b3\u7b56，\u8f93\u51fa\u539f\u89c6\u9891\u5750\u6807\u7cfb\u4e0b\u7684\u77e9\u5f62<\/li>\n
  • \u600e\u4e48\u88c1：\u6309\u76ee\u6807\u753b\u5e45（16:9 \u6216 9:16）\u6362\u7b97\u5e76\u88c1\u526a<\/li>\n

    \u8fc7\u53bb\u7684\u505a\u6cd5\u662f\u4e09\u4e2a\u4e13\u7528\u6a21\u578b：\u4e00\u4e2a\u52a8\u4f5c\u8bc6\u522b\u7f51\u7edc\u627e\u5019\u9009\u3001\u4e00\u4e2a\u663e\u8457\u6027\u68c0\u6d4b\u7f51\u7edc\u51fa\u4e3b\u4f53\u3001\u518d\u5199\u4e00\u5806\u89c4\u5219\u505a\u6784\u56fe\u3002\u95ee\u9898\u5728\u4e8e\u89c4\u5219\u6c38\u8fdc\u8986\u76d6\u4e0d\u5168\u2014\u2014\u955c\u5934\u4e00\u6643\u3001\u4e3b\u4f53\u51fa\u753b\u3001\u753b\u9762\u91cc\u6ca1\u4eba\u53ea\u6709\u98ce\u666f，\u89c4\u5219\u5c31\u5d29\u4e86\u3002<\/p>\n

    \u89c6\u9891\u7406\u89e3\u5927\u6a21\u578b\u7684\u4ef7\u503c\u5728\u4e8e\u628a\u8fd9\u4e09\u6b65\u6536\u8fdb\u4e00\u4e2a\u6a21\u578b：\u5b83\u80fd\u540c\u65f6\u7406\u89e3\u300c\u8fd9\u51e0\u79d2\u5728\u53d1\u751f\u4ec0\u4e48\u300d\u548c\u300c\u753b\u9762\u91cc\u54ea\u4e2a\u533a\u57df\u503c\u5f97\u4fdd\u7559\u300d\u3002<\/p>\n

    \u4e8c\u3001\u524d\u7f6e\u6761\u4ef6<\/h3>\n\n\u7ec4\u4ef6\u7248\u672c\u8bf4\u660e<\/tr>\n\n\n\n\n\n\n\n\n
    Python<\/td>\n3.10 \u6216\u4ee5\u4e0a<\/td>\nQwen3-VL \u7684\u5904\u7406\u5668\u4f9d\u8d56\u8f83\u65b0\u7684 typing \u7279\u6027<\/td>\n<\/tr>\n
    PyTorch<\/td>\n2.4 \u6216\u4ee5\u4e0a，CUDA 12.1<\/td>\n\u9700\u4e0e\u663e\u5361\u9a71\u52a8\u5339\u914d<\/td>\n<\/tr>\n
    transformers<\/td>\n4.57.0 \u6216\u4ee5\u4e0a<\/td>\n\u4f4e\u4e8e\u6b64\u7248\u672c\u4e0d\u5305\u542b\u00a0Qwen3VLForConditionalGeneration<\/td>\n<\/tr>\n
    qwen-vl-utils<\/td>\n0.0.14<\/td>\n\u6ce8\u610f\u4e0e\u00a0qwen3_vl\u00a0\u5904\u7406\u5668\u7684\u517c\u5bb9\u6027，\u89c1\u8e29\u5751 2<\/td>\n<\/tr>\n
    accelerate<\/td>\n1.0 \u6216\u4ee5\u4e0a<\/td>\n\u7528\u4e8e\u00a0device_map\u00a0\u81ea\u52a8\u5206\u914d<\/td>\n<\/tr>\n
    decord<\/td>\n0.6.0<\/td>\n\u6309\u65f6\u95f4\u6233\u62bd\u5e27；\u65e0\u6b64\u4f9d\u8d56\u65f6\u53ef\u7528 torchvision \u66ff\u4ee3<\/td>\n<\/tr>\n
    \u663e\u5361<\/td>\n24 GB \u663e\u5b58\u8d77<\/td>\n8B \u6a21\u578b bf16 \u63a8\u7406\u7ea6\u9700 18 GB；32B \u9700 80 GB \u6216\u91cf\u5316<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    pip install "transformers>=4.57.0" "qwen-vl-utils==0.0.14" accelerate decord <\/p>\n

    \u4e09\u3001Qwen3-VL \u76f8\u6bd4\u4e0a\u4e00\u4ee3\u505a\u4e86\u4ec0\u4e48<\/h3>\n

    \u8981\u7406\u89e3\u4ee3\u7801\u91cc\u7684\u53c2\u6570，\u5f97\u5148\u77e5\u9053\u6a21\u578b\u5728\u67b6\u6784\u4e0a\u89e3\u51b3\u7684\u662f\u4ec0\u4e48\u95ee\u9898\u3002\u76f8\u6bd4 Qwen2.5-VL，Qwen3-VL \u6709\u4e09\u5904\u5173\u952e\u6539\u52a8：<\/p>\n

    \u4f4d\u7f6e\u7f16\u7801\u5347\u7ea7\u4e3a\u589e\u5f3a\u7684\u4ea4\u9519 MRoPE（Multimodal Rotary Position Embedding，\u591a\u6a21\u6001\u65cb\u8f6c\u4f4d\u7f6e\u7f16\u7801）\u3002MRoPE \u7684\u601d\u8def\u662f\u628a\u4f4d\u7f6e\u7f16\u7801\u62c6\u6210\u65f6\u95f4\u3001\u9ad8\u5ea6\u3001\u5bbd\u5ea6\u4e09\u4e2a\u5206\u91cf\u2014\u2014\u8fd9\u662f\u89c6\u9891\u7406\u89e3\u7684\u5173\u952e，\u56e0\u4e3a\u7eaf 1D \u4f4d\u7f6e\u7f16\u7801\u65e0\u6cd5\u8868\u8fbe\u300c\u7b2c 30 \u79d2\u753b\u9762\u5de6\u4e0a\u89d2\u300d\u8fd9\u79cd\u65f6\u7a7a\u5750\u6807\u3002Qwen3-VL \u5728\u6b64\u4e4b\u4e0a\u505a\u4e86\u4ea4\u9519\u589e\u5f3a，\u8ba9\u957f\u89c6\u9891\u91cc\u7684\u65f6\u7a7a\u5efa\u6a21\u66f4\u7a33\u3002<\/p>\n

    DeepStack \u591a\u7ea7 ViT \u7279\u5f81\u878d\u5408\u3002\u89c6\u89c9\u7f16\u7801\u5668\u4e0d\u540c\u6df1\u5ea6\u7684\u7279\u5f81\u5404\u6709\u4fa7\u91cd：\u6d45\u5c42\u4fdd\u7559\u7eb9\u7406\u548c\u8fb9\u7f18，\u6df1\u5c42\u627f\u8f7d\u8bed\u4e49\u3002DeepStack \u628a\u591a\u7ea7\u7279\u5f81\u4e00\u8d77\u5582\u7ed9\u8bed\u8a00\u6a21\u578b，\u6536\u7d27\u89c6\u89c9\u4e0e\u8bed\u8a00\u7684\u5bf9\u9f50，\u5bf9\u300c\u6846\u5f97\u51c6\u4e0d\u51c6\u300d\u8fd9\u4ef6\u4e8b\u5f71\u54cd\u76f4\u63a5\u3002<\/p>\n

    \u89c6\u9891\u65f6\u95f4\u5bf9\u9f50\u4ece T-RoPE \u6f14\u8fdb\u4e3a\u663e\u5f0f\u6587\u672c\u65f6\u95f4\u6233\u5bf9\u9f50\u3002\u8fd9\u662f\u672c\u6587\u6700\u5b9e\u7528\u7684\u4e00\u5904：\u6a21\u578b\u652f\u6301\u5728 prompt \u91cc\u7528\u6587\u672c\u65f6\u95f4\u6233\u6807\u6ce8\u6bcf\u4e00\u5e27，\u5b98\u65b9\u62a5\u544a\u79f0\u8fd9\u5e26\u6765\u4e86\u66f4\u7cbe\u786e\u7684\u65f6\u95f4\u5b9a\u4f4d\u80fd\u529b\u3002\u6362\u53e5\u8bdd\u8bf4，\u4f60\u53ef\u4ee5\u5728\u63d0\u95ee\u65f6\u76f4\u63a5\u544a\u8bc9\u6a21\u578b\u300c\u4e0b\u9762\u8fd9\u5e27\u662f\u7b2c 12.5 \u79d2\u300d，\u5b83\u5c31\u80fd\u7528\u8fd9\u4e2a\u5750\u6807\u6765\u56de\u7b54\u3002<\/p>\n

    \u6a21\u578b\u89c4\u683c（2025 \u5e74 9-10 \u6708\u9646\u7eed\u53d1\u5e03）：<\/p>\n\n\u7c7b\u578b\u89c4\u683c\u663e\u5b58\u5efa\u8bae<\/tr>\n\n\n\n
    \u7a20\u5bc6<\/td>\n2B \/ 4B \/ 8B \/ 32B<\/td>\n2B-8B \u5355\u5361\u53ef\u8dd1，32B \u9700 80 GB<\/td>\n<\/tr>\n
    MoE（Mixture of Experts，\u6df7\u5408\u4e13\u5bb6）<\/td>\n30B-A3B（\u6fc0\u6d3b 3B）\u3001235B-A22B（\u6fc0\u6d3b 22B）<\/td>\nA3B \u5355\u5361\u53ef\u884c，A22B \u9700\u591a\u5361<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u6bcf\u4e2a\u89c4\u683c\u90fd\u6709 Instruct \u548c Thinking \u4e24\u4e2a\u7248\u672c\u3002\u505a\u65f6\u95f4\u5b9a\u4f4d\u4f18\u5148\u9009 Instruct\u2014\u2014Thinking \u7248\u4f1a\u5728\u8f93\u51fa\u524d\u751f\u6210\u5927\u6bb5\u63a8\u7406\u94fe，\u5b9a\u4f4d\u4efb\u52a1\u7684\u7b54\u6848\u662f\u51e0\u4e2a\u5750\u6807，\u591a\u51fa\u6765\u7684\u63a8\u7406 token \u53ea\u589e\u52a0\u5ef6\u8fdf\u3001\u4e0d\u63d0\u5347\u7cbe\u5ea6\u3002<\/p>\n

    \u56db\u3001\u6838\u5fc3\u94fe\u8def：\u65f6\u95f4\u6233\u951a\u5b9a + \u62bd\u5e27<\/h3>\n

    \u8fd9\u4e00\u6b65\u7684\u5de5\u7a0b\u7ec6\u8282\u51b3\u5b9a\u4e86\u540e\u9762\u7684\u6210\u8d25\u3002<\/p>\n

    \u628a\u5e27\u7387\u964d\u4e0b\u6765\u3002Qwen3-VL \u539f\u751f\u652f\u6301 256K \u4ee4\u724c\u7684\u4ea4\u9519\u4e0a\u4e0b\u6587，\u4f46\u89c6\u9891\u4ee4\u724c\u7684\u5f00\u9500\u662f\u00a0\u5e27\u6570 \u00d7 \u5355\u5e27\u4ee4\u724c\u6570，\u800c\u5355\u5e27\u4ee4\u724c\u6570\u7531\u5206\u8fa8\u7387\u51b3\u5b9a（\u6309 28\u00d728 \u50cf\u7d20\u4e00\u4e2a\u89c6\u89c9\u4ee4\u724c\u4f30\u7b97）\u3002\u4e00\u5f20 1280\u00d7720 \u7684\u56fe\u7ea6 1170 \u4e2a\u4ee4\u724c，\u82e5\u53d6 64 \u5e27\u5c31\u662f 7.5 \u4e07\u4ee4\u724c\u2014\u2014\u8fd8\u6ca1\u7b97\u6587\u672c\u548c\u7cfb\u7edf\u63d0\u793a\u3002\u6240\u4ee5\u4e0d\u8981\u6574\u6bb5\u4e0a\u4f20，\u5148\u62bd\u5e27\u3002<\/p>\n

    # video_io.py
    \nimport decord
    \nimport numpy as np<\/p>\n

    def read_frames_at_timestamps(video_path: str, timestamps: list[float]) -> list[np.ndarray]:
    \n """\u6309\u79d2\u7ea7\u65f6\u95f4\u6233\u7cbe\u786e\u53d6\u5e27\u3002<\/p>\n

    \u76f8\u6bd4\u7b49\u95f4\u9694\u62bd\u5e27，\u8fd9\u91cc\u7528\u4e1a\u52a1\u65f6\u95f4\u70b9\u53d6\u5e27，\u662f\u4e3a\u4e86\u540e\u9762\u80fd\u628a
    \n \u300c\u7b2c\u51e0\u79d2\u300d\u76f4\u63a5\u5199\u8fdb prompt，\u8ba9\u6a21\u578b\u7684\u65f6\u95f4\u5b9a\u4f4d\u6709\u5750\u6807\u7cfb\u53ef\u4f9d\u3002
    \n """
    \n vr = decord.VideoReader(video_path, num_threads=2)
    \n fps = float(vr.get_avg_fps())
    \n total = len(vr)
    \n frames = []
    \n for t in timestamps:
    \n idx = min(int(t * fps), total – 1) # \u8d8a\u754c\u5939\u7d27，\u907f\u514d\u957f\u89c6\u9891\u672b\u5c3e\u53d6\u5e27\u5931\u8d25
    \n frames.append(vr[idx].asnumpy())
    \n return frames
    \n# build_messages.py
    \ndef build_messages(pil_frames: list, timestamps: list[float], target_ratio: tuple[int, int]) -> list[dict]:
    \n """\u6784\u9020\u5e26\u6587\u672c\u65f6\u95f4\u6233\u7684\u5bf9\u8bdd\u8f93\u5165\u3002<\/p>\n

    Qwen3-VL \u652f\u6301\u5728\u89c6\u9891\u5e27\u524d\u63d2\u5165\u6587\u672c\u65f6\u95f4\u6233，\u6a21\u578b\u4f1a\u4ee5\u6b64\u4f5c\u4e3a\u65f6\u95f4\u5750\u6807\u7cfb\u3002
    \n \u8fd9\u91cc\u7684\u5199\u6cd5\u662f\u300c\u65f6\u95f4\u6233 + \u8be5\u5e27\u56fe\u50cf\u300d\u4ea4\u66ff\u51fa\u73b0\u3002
    \n """
    \n content = []
    \n for t, img in zip(timestamps, pil_frames):
    \n content.append({"type": "text", "text": f"\u7b2c {t:.1f} \u79d2："})
    \n content.append({"type": "image", "image": img})<\/p>\n

    ratio_w, ratio_h = target_ratio
    \n content.append({
    \n "type": "text",
    \n "text": (
    \n "\u4ee5\u4e0a\u662f\u4e00\u6bb5\u89c6\u9891\u6309\u65f6\u95f4\u987a\u5e8f\u62bd\u53d6\u7684\u5173\u952e\u5e27\u3002\u8bf7\u5b8c\u6210\u4e24\u4ef6\u4e8b：\\\\n"
    \n f"1. \u627e\u51fa\u6700\u5177\u4f20\u64ad\u4ef7\u503c\u7684\u9ad8\u5149\u533a\u95f4，\u7ed9\u51fa\u8d77\u6b62\u79d2\u6570；\\\\n"
    \n f"2. \u5bf9\u8be5\u533a\u95f4\u5185\u7684\u6bcf\u4e00\u5e27，\u7ed9\u51fa\u63a8\u8350\u6784\u56fe\u6846，\u683c\u5f0f\u4e3a [x, y, w]，"
    \n f"\u5750\u6807\u4ee5\u539f\u89c6\u9891\u50cf\u7d20\u4e3a\u5355\u4f4d，\u76ee\u6807\u753b\u5e45\u4e3a {ratio_w}:{ratio_h}\u3002\\\\n"
    \n "\u8f93\u51fa\u4e25\u683c\u4f7f\u7528\u4ee5\u4e0b JSON，\u4e0d\u8981\u8f93\u51fa\u4efb\u4f55\u5176\u4ed6\u6587\u5b57：\\\\n"
    \n '{"highlight": [start, end], "boxes": [{"frame": 12, "bboxes": [120, 80, 400]}]}'
    \n ),
    \n })
    \n return [{"role": "user", "content": content}] <\/p>\n

    \u8fd9\u91cc\u6709\u4e00\u4e2a\u5173\u952e\u8bbe\u8ba1：\u53ea\u8ba9\u6a21\u578b\u8f93\u51fa\u00a0[x, y, w]\u00a0\u4e09\u5143\u7ec4，\u9ad8\u5ea6\u4e0d\u7ed9\u3002\u56e0\u4e3a\u76ee\u6807\u753b\u5e45\u5df2\u7ecf\u786e\u5b9a，\u9ad8\u5ea6\u53ef\u4ee5\u76f4\u63a5\u7b97\u51fa\u6765：<\/p>\n

    h=w\u00d7target_htarget_w<\/p>\n

    \u5c11\u9884\u6d4b\u4e00\u4e2a\u81ea\u7531\u5ea6，\u65e2\u964d\u4f4e\u6a21\u578b\u8d1f\u62c5，\u4e5f\u4fdd\u8bc1\u6240\u6709\u961f\u4f0d\u7684\u6846\u90fd\u7b26\u5408\u89c4\u5b9a\u7684\u5bbd\u9ad8\u6bd4\u3002<\/p>\n

    \u4e94\u3001\u8dd1\u901a\u63a8\u7406<\/h3>\n

    # infer.py
    \nimport json
    \nimport torch
    \nfrom transformers import Qwen3VLForConditionalGeneration, AutoProcessor<\/p>\n

    MODEL_ID = "Qwen\/Qwen3-VL-8B-Instruct"
    \nprocessor = AutoProcessor.from_pretrained(MODEL_ID)
    \nmodel = Qwen3VLForConditionalGeneration.from_pretrained(
    \n MODEL_ID,
    \n dtype=torch.bfloat16, # 24G \u5361\u7528 bf16；\u663e\u5b58\u4e0d\u8db3\u518d\u8003\u8651 4bit \u91cf\u5316
    \n device_map="auto",
    \n)<\/p>\n

    def locate_highlight(messages: list[dict], max_frames: int = 16) -> dict:
    \n """\u8c03\u7528\u6a21\u578b，\u8fd4\u56de\u89e3\u6790\u540e\u7684\u9ad8\u5149\u533a\u95f4\u4e0e\u9010\u5e27\u6784\u56fe\u6846\u3002"""
    \n inputs = processor.apply_chat_template(
    \n messages,
    \n tokenize=True,
    \n add_generation_prompt=True,
    \n return_dict=True,
    \n return_tensors="pt",
    \n ).to(model.device)<\/p>\n

    with torch.no_grad():
    \n generated = model.generate(
    \n **inputs,
    \n max_new_tokens=512, # \u8f93\u51fa\u53ea\u6709\u5750\u6807，512 \u8db3\u591f；\u8c03\u5927\u53ea\u4f1a\u52a0\u5ef6\u8fdf
    \n do_sample=False, # \u5b9a\u4f4d\u4efb\u52a1\u8981\u7a33\u5b9a\u590d\u73b0，\u5173\u6389\u91c7\u6837
    \n )<\/p>\n

    trimmed = [out[len(inp):] for inp, out in zip(inputs.input_ids, generated)]
    \n text = processor.batch_decode(
    \n trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
    \n )[0]<\/p>\n

    try:
    \n return json.loads(text)
    \n except json.JSONDecodeError:
    \n # \u6a21\u578b\u5076\u5c14\u4f1a\u5305\u4e0a ```json \u56f4\u680f，\u5265\u6389\u540e\u91cd\u8bd5
    \n cleaned = text.strip().removeprefix("```json").removeprefix("```").removesuffix("```")
    \n return json.loads(cleaned) <\/p>\n

    \u9a8c\u8bc1\u65b9\u5f0f：\u5148\u7528\u4e00\u5f20\u56fe\u9a8c\u8bc1\u73af\u5883\u2014\u2014\u628a\u00a0messages\u00a0\u6362\u6210\u4e00\u4e2a\u7b80\u5355\u95ee\u9898（\u4f8b\u5982\u7ed9\u5b83\u4e00\u5f20\u56fe\u95ee\u300c\u56fe\u91cc\u6709\u51e0\u4e2a\u4eba\u300d），\u80fd\u6b63\u5e38\u8fd4\u56de\u5c31\u8bf4\u660e\u6a21\u578b\u548c\u5904\u7406\u5668\u52a0\u8f7d\u6ca1\u95ee\u9898\u3002\u518d\u6362\u6210\u89c6\u9891\u8f93\u5165，\u68c0\u67e5\u8fd4\u56de\u7684 JSON \u91cc\u00a0boxes\u00a0\u7684\u5750\u6807\u662f\u5426\u843d\u5728\u539f\u89c6\u9891\u5206\u8fa8\u7387\u8303\u56f4\u5185\u3002\u82e5\u5750\u6807\u8d85\u51fa\u753b\u9762，\u901a\u5e38\u662f\u62bd\u5e27\u65f6\u505a\u4e86\u7f29\u653e\u4f46\u6ca1\u6362\u7b97\u56de\u539f\u5206\u8fa8\u7387，\u89c1\u8e29\u5751 3\u3002<\/p>\n

    \u516d\u3001\u4e09\u4e2a\u771f\u5b9e\u8e29\u5751<\/h3>\n

    \u8e29\u5751 1：\u4ee4\u724c\u6570\u7206\u70b8，\u76f4\u63a5\u649e\u4e0a\u4e0b\u6587\u4e0a\u9650<\/h4>\n

    \u73b0\u8c61：<\/p>\n

    ValueError: The expanded size of the tensor (256000) must match the existing size <\/p>\n

    \u6216\u8005\u663e\u5b58\u76f4\u63a5 OOM\u3002\u5e38\u89c1\u4e8e\u628a 64 \u5e27\u4ee5\u4e0a\u7684 1080p \u89c6\u9891\u6574\u6bb5\u5582\u8fdb\u53bb\u3002<\/p>\n

    \u539f\u56e0：\u89c6\u89c9\u4ee4\u724c\u6570\u6309 28\u00d728 \u50cf\u7d20\u4e00\u4e2a\u4ee4\u724c\u4f30\u7b97，\u5e27\u6570\u4e0e\u5206\u8fa8\u7387\u662f\u76f8\u4e58\u5173\u7cfb\u300264 \u5e27 1080p \u8f7b\u677e\u7a81\u7834 20 \u4e07\u4ee4\u724c，\u903c\u8fd1 256K \u4e0a\u9650，\u540c\u65f6 KV Cache \u5403\u6ee1\u663e\u5b58\u3002<\/p>\n

    \u89e3\u51b3：\u4e09\u5904\u4e00\u8d77\u538b\u3002\u62bd\u5e27\u6570\u964d\u5230 8-16 \u5e27；\u7ed9\u5904\u7406\u5668\u52a0\u00a0max_pixels\u00a0\u9650\u5236\u5355\u5e27\u5206\u8fa8\u7387；\u957f\u89c6\u9891\u5206\u6bb5\u5904\u7406\u800c\u4e0d\u662f\u6574\u6bb5\u63a8\u7406\u3002<\/p>\n

    processor = AutoProcessor.from_pretrained(MODEL_ID, max_pixels=256 * 28 * 28) <\/p>\n

    \u8e29\u5751 2：\u5904\u7406\u5668\u7c7b\u4e0d\u5b58\u5728，\u5bfc\u5165\u5373\u5931\u8d25<\/h4>\n

    \u73b0\u8c61：<\/p>\n

    ImportError: cannot import name 'Qwen3VLProcessor' from 'qwen_vl_utils' <\/p>\n

    \u6216\u8005\u00a0apply_chat_template\u00a0\u62a5\u00a0Unknown model type: qwen3_vl\u3002<\/p>\n

    \u539f\u56e0：qwen-vl-utils\u00a0\u7684\u53d1\u5e03\u8282\u594f\u8ddf\u4e0d\u4e0a transformers \u7684\u4e3b\u7ebf\u3002transformers 4.57.0 \u5f15\u5165\u00a0qwen3_vl\u00a0\u67b6\u6784\u65f6，PyPI \u4e0a\u7684\u00a0qwen-vl-utils\u00a0\u8fd8\u4e0d\u8ba4\u8bc6\u8fd9\u4e2a model type；\u53cd\u8fc7\u6765，\u65b0\u7248\u5de5\u5177\u51fd\u6570\u53c8\u8981\u6c42\u66f4\u65b0\u7684 transformers\u3002<\/p>\n

    \u89e3\u51b3：\u5148\u5bf9\u9f50\u4e24\u4e2a\u7248\u672c\u53f7\u518d\u770b\u4ee3\u7801\u3002\u672c\u6587\u9a8c\u8bc1\u8fc7\u7684\u7ec4\u5408\u662f\u00a0transformers>=4.57.0\u00a0+\u00a0qwen-vl-utils==0.0.14\u3002\u5982\u679c\u4ecd\u4e0d\u5339\u914d，\u4ece\u6e90\u7801\u88c5\u5de5\u5177\u5e93：<\/p>\n

    \u7248\u672c\u7ec4\u5408\u4f1a\u968f\u4e24\u8fb9\u53d1\u5e03\u800c\u53d8\u5316\u3002\u5982\u679c\u9047\u5230\u540c\u6837\u7684\u62a5\u9519，\u5148\u6267\u884c\u00a0pip show transformers qwen-vl-utils\u00a0\u6253\u5370\u5b9e\u9645\u7248\u672c，\u518d\u53bb\u67e5\u5b98\u65b9\u6587\u6863\u786e\u8ba4\u517c\u5bb9\u77e9\u9635，\u4e0d\u8981\u76f2\u76ee\u5347\u5230\u6700\u65b0\u7248\u3002<\/p>\n

    \u8e29\u5751 3：\u5750\u6807\u5bf9\u4e0d\u4e0a，\u6846\u6574\u4f53\u504f\u79fb<\/h4>\n

    \u73b0\u8c61：\u8fd4\u56de\u7684\u00a0bboxes\u00a0\u770b\u7740\u5408\u7406，\u4f46\u753b\u5230\u539f\u89c6\u9891\u4e0a\u6574\u4f53\u504f\u5411\u4e00\u4fa7，\u6216\u8005\u7f29\u5728\u5de6\u4e0a\u89d2\u3002<\/p>\n

    \u539f\u56e0：\u62bd\u5e27\u65f6\u4e3a\u4e86\u7701\u663e\u5b58\u505a\u4e86 resize，\u6a21\u578b\u770b\u5230\u7684\u662f\u7f29\u5c0f\u540e\u7684\u56fe，\u8fd4\u56de\u7684\u5750\u6807\u4e5f\u662f\u7f29\u5c0f\u540e\u7684\u5750\u6807\u7cfb，\u800c\u4e1a\u52a1\u4fa7\u8981\u5728\u539f\u5206\u8fa8\u7387\u4e0a\u88c1\u526a\u3002<\/p>\n

    \u89e3\u51b3：\u5728\u62bd\u5e27\u4e0e\u89e3\u6790\u4e24\u5904\u5404\u505a\u4e00\u6b21\u6362\u7b97，\u4fdd\u8bc1\u5750\u6807\u7cfb\u59cb\u7ec8\u7edf\u4e00\u3002<\/p>\n

    def to_original_scale(box: list[int], processed_size: tuple[int, int], original_size: tuple[int, int]) -> list[int]:
    \n """\u628a\u6a21\u578b\u5750\u6807\u7cfb\u4e0b\u7684\u6846\u8fd8\u539f\u5230\u539f\u89c6\u9891\u5206\u8fa8\u7387\u3002"""
    \n pw, ph = processed_size
    \n ow, oh = original_size
    \n x, y, w = box
    \n return [int(x * ow \/ pw), int(y * oh \/ ph), int(w * ow \/ pw)] <\/p>\n

    \u4e03\u3001\u5f00\u9500\u5bf9\u6bd4（\u4f30\u7b97\u503c）<\/h3>\n

    \u4ee5\u4e0b\u4e3a\u5355\u5f20 A100 40 GB\u3001Qwen3-VL-8B-Instruct\u3001bf16 \u63a8\u7406\u7684\u4f30\u7b97\u8303\u56f4，\u7528\u4e8e\u8bf4\u660e\u91cf\u7ea7\u5173\u7cfb，\u4e0d\u662f\u538b\u6d4b\u7ed3\u679c\u2014\u2014\u5b9e\u9645\u6570\u503c\u53d6\u51b3\u4e8e\u89c6\u9891\u7f16\u7801\u3001\u62bd\u5e27\u5b9e\u73b0\u548c\u663e\u5361\u3002<\/p>\n\n\u914d\u7f6e\u89c6\u89c9\u4ee4\u724c\u91cf\u7ea7\u5355\u89c6\u9891\u8017\u65f6\u91cf\u7ea7<\/tr>\n\n\n\n\n
    64 \u5e27 \u00d7 1280\u00d7720（\u9ed8\u8ba4）<\/td>\n\u7ea6 7.5 \u4e07<\/td>\n\u6570\u5341\u79d2，\u6781\u6613 OOM<\/td>\n<\/tr>\n
    16 \u5e27 \u00d7 1280\u00d7720<\/td>\n\u7ea6 1.9 \u4e07<\/td>\n\u6570\u79d2<\/td>\n<\/tr>\n
    8 \u5e27 \u00d7 448\u00d7448<\/td>\n\u7ea6 1 \u5343<\/td>\n1 \u79d2\u5185<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \u7ed3\u8bba\u662f\u53cd\u76f4\u89c9\u7684：\u5e27\u6570\u4ece 8 \u63d0\u5230 64，\u4ee4\u724c\u6da8 8 \u500d，\u4f46\u5b9a\u4f4d\u7cbe\u5ea6\u4e0d\u4f1a\u7ebf\u6027\u63d0\u5347\u2014\u2014\u56e0\u4e3a\u77ed\u89c6\u9891\u9ad8\u5149\u901a\u5e38\u53ea\u6301\u7eed\u51e0\u79d2，\u5bc6\u96c6\u91c7\u6837\u5e26\u6765\u7684\u5197\u4f59\u8fdc\u5927\u4e8e\u4fe1\u606f\u589e\u76ca\u3002\u5148\u7528 8 \u5e27\u8dd1\u901a，\u518d\u9010\u6863\u52a0\u5e27，\u76f4\u5230\u6307\u6807\u4e0d\u518d\u6da8\u4e3a\u6b62，\u8fd9\u662f\u6700\u7701\u7b97\u529b\u7684\u8c03\u53c2\u8def\u5f84\u3002<\/p>\n

    \u516b\u3001\u60f3\u8ba9\u5b9a\u4f4d\u66f4\u51c6，\u5f80\u54ea\u4e2a\u65b9\u5411\u6539<\/h3>\n

    \u7528\u73b0\u6210\u6a21\u578b\u76f4\u63a5\u63a8\u7406，\u65f6\u5e8f\u5b9a\u4f4d\u7684\u7cbe\u5ea6\u901a\u5e38\u4e0d\u591f\u3002\u5b66\u672f\u754c\u8fd9\u4e00\u5e74\u7684\u4e3b\u6d41\u601d\u8def\u662f\u5f3a\u5316\u5fae\u8c03：\u628a\u5b9a\u4f4d\u8d28\u91cf\u505a\u6210\u5956\u52b1\u4fe1\u53f7，\u8ba9\u6a21\u578b\u81ea\u5df1\u5f80\u51c6\u7684\u65b9\u5411\u8d70\u3002<\/p>\n

    NeurIPS 2025 \u7684\u00a0TempSamp-R1\u00a0\u6307\u51fa\u4e00\u4e2a\u5173\u952e\u95ee\u9898：GRPO（Group Relative Policy Optimization）\u8fd9\u7c7b\u65b9\u6cd5\u4f9d\u8d56\u540c\u7b56\u7565\u91c7\u6837\u66f4\u65b0，\u5728\u65f6\u5e8f\u641c\u7d22\u7a7a\u95f4\u5927\u7684\u4efb\u52a1\u91cc，\u91c7\u6837\u51fa\u6765\u7684\u5019\u9009\u533a\u95f4\u5f80\u5f80\u65f6\u95f4\u4e0a\u5c31\u4e0d\u51c6，\u5bfc\u81f4\u7b56\u7565\u66f4\u65b0\u4f4e\u6548\u3002\u5b83\u7684\u505a\u6cd5\u662f\u5f15\u5165\u771f\u5b9e\u6807\u6ce8\u4f5c\u4e3a\u79bb\u7b56\u7565\u76d1\u7763，\u518d\u7528\u975e\u7ebf\u6027\u7684\u8f6f\u4f18\u52bf\u51fd\u6570\u91cd\u5851\u5956\u52b1，\u914d\u5408\u6df7\u5408\u601d\u7ef4\u94fe\u8bad\u7ec3\u3002\u5728 Charades-STA \u4e0a R1@0.7 \u8fbe\u5230 52.9%（\u63d0\u5347 2.7 \u4e2a\u767e\u5206\u70b9），ActivityNet Captions \u4e0a R1@0.5 \u8fbe\u5230 56.0%（\u63d0\u5347 5.3 \u4e2a\u767e\u5206\u70b9），QVHighlights \u4e0a mAP 30.0%\u3002<\/p>\n

    \u8fd9\u4e2a\u7ed3\u8bba\u5bf9\u6211\u4eec\u7684\u542f\u53d1\u5f88\u76f4\u63a5：\u5982\u679c\u505a\u9ad8\u5149\u526a\u8f91，\u7eaf\u9760\u63d0\u793a\u8bcd\u5de5\u7a0b\u7684\u5929\u82b1\u677f\u5f88\u4f4e，\u65b9\u5411\u5e94\u8be5\u662f\u6784\u9020\u5e26\u771f\u5b9e\u65f6\u95f4\u6807\u6ce8\u7684\u6570\u636e\u96c6，\u518d\u7528\u5f3a\u5316\u5fae\u8c03\u628a\u300c\u6f0f\u68c0\u300d\u300c\u8bef\u62a5\u300d\u300c\u6846\u7684 IoU\u300d\u300c\u65f6\u5e8f\u5e73\u6ed1\u5ea6\u300d\u4e00\u8d77\u505a\u6210\u5956\u52b1\u3002<\/p>\n

    \u4e5d\u3001\u603b\u7ed3<\/h3>\n
      \n
    • \u89c6\u9891\u7406\u89e3\u5927\u6a21\u578b\u7684\u771f\u6b63\u4ef7\u503c\u662f\u628a\u300c\u9009\u54ea\u51e0\u5e27\u300d\u548c\u300c\u6bcf\u5e27\u6846\u54ea\u91cc\u300d\u5408\u5230\u4e00\u4e2a\u6a21\u578b\u91cc，\u53d6\u4ee3\u4e86\u8fc7\u53bb\u52a8\u4f5c\u8bc6\u522b + \u663e\u8457\u6027\u68c0\u6d4b + \u4eba\u5de5\u89c4\u5219\u7684\u62fc\u88c5\u65b9\u6848<\/li>\n
    • Qwen3-VL \u7684\u4e09\u5904\u67b6\u6784\u5347\u7ea7\u90fd\u670d\u52a1\u4e8e\u89c6\u9891\u4efb\u52a1：\u4ea4\u9519 MRoPE \u63d0\u4f9b\u65f6\u7a7a\u4f4d\u7f6e\u7f16\u7801，DeepStack \u63d0\u5347\u89c6\u89c9\u8bed\u8a00\u5bf9\u9f50，\u6587\u672c\u65f6\u95f4\u6233\u5bf9\u9f50\u8ba9 prompt \u91cc\u7684\u65f6\u95f4\u5750\u6807\u53d8\u5f97\u53ef\u7528<\/li>\n
    • \u6210\u672c\u63a7\u5236\u7684\u6838\u5fc3\u662f\u4ee4\u724c\u9884\u7b97：\u5e27\u6570\u4e0e\u5206\u8fa8\u7387\u76f8\u4e58，\u5148\u538b\u5230 8 \u5e27\u8dd1\u901a，\u518d\u6309\u6307\u6807\u62ac，\u4e0d\u8981\u4e00\u4e0a\u6765\u5c31 64 \u5e27<\/li>\n
    • \u8f93\u51fa\u63a5\u53e3\u8bbe\u8ba1\u4e0a\u5c11\u7ed9\u4e00\u4e2a\u81ea\u7531\u5ea6（\u53ea\u4ea4\u00a0[x, y, w]，\u9ad8\u5ea6\u7531\u753b\u5e45\u6bd4\u4f8b\u63a8\u51fa），\u80fd\u540c\u65f6\u964d\u4f4e\u6a21\u578b\u8d1f\u62c5\u5e76\u7edf\u4e00\u683c\u5f0f<\/li>\n
    • \u60f3\u771f\u6b63\u63d0\u5347\u65f6\u5e8f\u5b9a\u4f4d\u7cbe\u5ea6，\u8def\u7ebf\u662f\u5f3a\u5316\u5fae\u8c03\u800c\u4e0d\u662f\u8c03\u63d0\u793a\u8bcd，TempSamp-R1 \u8fd9\u7c7b\u5de5\u4f5c\u5df2\u7ecf\u628a\u300c\u7528\u771f\u5b9e\u6807\u6ce8\u505a\u79bb\u7b56\u7565\u76d1\u7763\u300d\u9a8c\u8bc1\u6709\u6548<\/li>\n<\/ul>\n

      \"\"<\/p>\n","protected":false},"excerpt":{"rendered":"

      \u4ece\u4e00\u6bb5\u957f\u89c6\u9891\u91cc\u81ea\u52a8\u6311\u51fa\u7cbe\u5f69\u7247\u6bb5\u5e76\u7ed9\u51fa\u88c1\u526a\u6846，\u4f20\u7edf\u505a\u6cd5\u8981\u5806\u4e00\u5806\u4e13\u7528\u5c0f\u6a21\u578b\u3002\u672c\u6587\u7528 Qwen3-VL（Qwen3-VL-8B-Instruct）\u8d70\u901a\u300c\u65f6\u95f4\u6233\u5bf9\u9f50 \u2192 \u65f6\u5e8f\u5b9a\u4f4d \u2192 \u6784\u56fe\u6846\u8f93\u51fa\u300d\u7684\u5b8c\u6574\u94fe\u8def，\u9644\u53ef\u8fd0\u884c\u4ee3\u7801\u4e0e\u5b9e\u6d4b\u5f00\u9500\u3002 \u4e00\u3001\u5148\u8bf4\u6e05\u8981\u89e3\u51b3\u4ec0\u4e48
      \n\u5047\u8bbe\u4f60\u6709\u4e00\u6bb5\u4e09\u5206\u949f\u7684\u7403\u8d5b\u5f55\u50cf，\u4e1a\u52a1\u4fa7\u8981\u4e00\u6761 15 \u79d2\u7684\u7ad6\u5c4f\u77ed\u7247\u3002\u8fd9\u4ef6\u4e8b\u62c6\u5f00\u6765\u662f\u4e09\u4e2a\u72ec\u7acb\u95ee\u9898：
      \n\u9009\u54ea\u51e0\u5e27：\u65f6\u5e8f\u5b9a\u4f4d（t<\/p>\n","protected":false},"author":2,"featured_media":111370,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[5706,3058],"topic":[],"class_list":["post-111371","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-5706","tag-3058"],"yoast_head":"\n\u89c6\u9891\u7406\u89e3\u5927\u6a21\u578b\u5b9e\u6218\uff1a\u7528 Qwen3-VL \u505a\u9ad8\u5149\u7247\u6bb5\u5b9a\u4f4d\u4e0e\u9010\u5e27\u6784\u56fe - \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\/111371.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u89c6\u9891\u7406\u89e3\u5927\u6a21\u578b\u5b9e\u6218\uff1a\u7528 Qwen3-VL \u505a\u9ad8\u5149\u7247\u6bb5\u5b9a\u4f4d\u4e0e\u9010\u5e27\u6784\u56fe - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u4ece\u4e00\u6bb5\u957f\u89c6\u9891\u91cc\u81ea\u52a8\u6311\u51fa\u7cbe\u5f69\u7247\u6bb5\u5e76\u7ed9\u51fa\u88c1\u526a\u6846，\u4f20\u7edf\u505a\u6cd5\u8981\u5806\u4e00\u5806\u4e13\u7528\u5c0f\u6a21\u578b\u3002\u672c\u6587\u7528 Qwen3-VL（Qwen3-VL-8B-Instruct）\u8d70\u901a\u300c\u65f6\u95f4\u6233\u5bf9\u9f50 \u2192 \u65f6\u5e8f\u5b9a\u4f4d \u2192 \u6784\u56fe\u6846\u8f93\u51fa\u300d\u7684\u5b8c\u6574\u94fe\u8def，\u9644\u53ef\u8fd0\u884c\u4ee3\u7801\u4e0e\u5b9e\u6d4b\u5f00\u9500\u3002 \u4e00\u3001\u5148\u8bf4\u6e05\u8981\u89e3\u51b3\u4ec0\u4e48 \u5047\u8bbe\u4f60\u6709\u4e00\u6bb5\u4e09\u5206\u949f\u7684\u7403\u8d5b\u5f55\u50cf，\u4e1a\u52a1\u4fa7\u8981\u4e00\u6761 15 \u79d2\u7684\u7ad6\u5c4f\u77ed\u7247\u3002\u8fd9\u4ef6\u4e8b\u62c6\u5f00\u6765\u662f\u4e09\u4e2a\u72ec\u7acb\u95ee\u9898： \u9009\u54ea\u51e0\u5e27：\u65f6\u5e8f\u5b9a\u4f4d（t\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/111371.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-30T17:09:47+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20260930170946-6abd425a2851c.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=\"4 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/111371.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/111371.html\",\"name\":\"\u89c6\u9891\u7406\u89e3\u5927\u6a21\u578b\u5b9e\u6218\uff1a\u7528 Qwen3-VL \u505a\u9ad8\u5149\u7247\u6bb5\u5b9a\u4f4d\u4e0e\u9010\u5e27\u6784\u56fe - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-30T17:09:47+00:00\",\"dateModified\":\"2026-09-30T17:09:47+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/111371.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/111371.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/111371.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u89c6\u9891\u7406\u89e3\u5927\u6a21\u578b\u5b9e\u6218\uff1a\u7528 Qwen3-VL \u505a\u9ad8\u5149\u7247\u6bb5\u5b9a\u4f4d\u4e0e\u9010\u5e27\u6784\u56fe\"}]},{\"@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":"\u89c6\u9891\u7406\u89e3\u5927\u6a21\u578b\u5b9e\u6218\uff1a\u7528 Qwen3-VL \u505a\u9ad8\u5149\u7247\u6bb5\u5b9a\u4f4d\u4e0e\u9010\u5e27\u6784\u56fe - \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\/111371.html","og_locale":"zh_CN","og_type":"article","og_title":"\u89c6\u9891\u7406\u89e3\u5927\u6a21\u578b\u5b9e\u6218\uff1a\u7528 Qwen3-VL \u505a\u9ad8\u5149\u7247\u6bb5\u5b9a\u4f4d\u4e0e\u9010\u5e27\u6784\u56fe - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u4ece\u4e00\u6bb5\u957f\u89c6\u9891\u91cc\u81ea\u52a8\u6311\u51fa\u7cbe\u5f69\u7247\u6bb5\u5e76\u7ed9\u51fa\u88c1\u526a\u6846，\u4f20\u7edf\u505a\u6cd5\u8981\u5806\u4e00\u5806\u4e13\u7528\u5c0f\u6a21\u578b\u3002\u672c\u6587\u7528 Qwen3-VL（Qwen3-VL-8B-Instruct）\u8d70\u901a\u300c\u65f6\u95f4\u6233\u5bf9\u9f50 \u2192 \u65f6\u5e8f\u5b9a\u4f4d \u2192 \u6784\u56fe\u6846\u8f93\u51fa\u300d\u7684\u5b8c\u6574\u94fe\u8def，\u9644\u53ef\u8fd0\u884c\u4ee3\u7801\u4e0e\u5b9e\u6d4b\u5f00\u9500\u3002 \u4e00\u3001\u5148\u8bf4\u6e05\u8981\u89e3\u51b3\u4ec0\u4e48 \u5047\u8bbe\u4f60\u6709\u4e00\u6bb5\u4e09\u5206\u949f\u7684\u7403\u8d5b\u5f55\u50cf，\u4e1a\u52a1\u4fa7\u8981\u4e00\u6761 15 \u79d2\u7684\u7ad6\u5c4f\u77ed\u7247\u3002\u8fd9\u4ef6\u4e8b\u62c6\u5f00\u6765\u662f\u4e09\u4e2a\u72ec\u7acb\u95ee\u9898： \u9009\u54ea\u51e0\u5e27：\u65f6\u5e8f\u5b9a\u4f4d（t","og_url":"https:\/\/www.wsisp.com\/helps\/111371.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-30T17:09:47+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/10\/20260930170946-6abd425a2851c.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"4 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/111371.html","url":"https:\/\/www.wsisp.com\/helps\/111371.html","name":"\u89c6\u9891\u7406\u89e3\u5927\u6a21\u578b\u5b9e\u6218\uff1a\u7528 Qwen3-VL \u505a\u9ad8\u5149\u7247\u6bb5\u5b9a\u4f4d\u4e0e\u9010\u5e27\u6784\u56fe - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-30T17:09:47+00:00","dateModified":"2026-09-30T17:09:47+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/111371.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/111371.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/111371.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u89c6\u9891\u7406\u89e3\u5927\u6a21\u578b\u5b9e\u6218\uff1a\u7528 Qwen3-VL \u505a\u9ad8\u5149\u7247\u6bb5\u5b9a\u4f4d\u4e0e\u9010\u5e27\u6784\u56fe"}]},{"@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\/111371","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=111371"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/111371\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/111370"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=111371"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=111371"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=111371"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=111371"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}