燃
We recommend passing the video URL and controlling the sampling frequency via the fps parameter. For details, please refer to: https://github.com/QwenLM/Qwen3-VL/pull/1644.
We have recently updated the documentation for video requests. You can take a look. @LinaZhangCoding If your issue has been resolved, please close the issue.
> Hi, [@nutsintheshell](https://github.com/nutsintheshell) , > > I have conducted the evaluation but also observed a score close to 0 (0.001, actually). > > [@wulipc](https://github.com/wulipc) , would you please help check...
@1994 Hi, thank you for your interest in Qwen! In fact, the image processing pipeline is vLLM → transformers → torchvision.transforms.resize. If you're launching an online server, it does not...
@Cristhine 感谢对 Qwen 的关注，下面文档中提供了 Benchmark 的方式，您可以参加运行下： https://docs.vllm.ai/projects/recipes/en/latest/Qwen/Qwen3-VL.html 我在 H100 的机器上进行了测试，当 `request-rate` 设置为 10 的时候，输出如下，可以看到首包平均时间在 434.92 ms 左右； 注意 `request-rate` 的值会严重影响 TTFT， 如果设置过大，内部对产生排队、调度等额外的开销。 ``` ============ Serving Benchmark Result ============ Successful requests:...
This is normal; just wait a moment.
数据只是一方面，更多信息请留意我们未来发布的技术报告。
@yss0729 16G显存加载 7B 模型确实有点困难，也不是安装 flash attention 能解决的；如 @JJJYmmm 所说，可以尝试 7B量化版本/3b的model； 或者你有多卡的话，可以尝试 开 TP。
Please check your installation of transformers. Good luck.
@zjx-ERROR 请提供一下原始视频，我复现下问题： 可以发送至：[email protected]