timelens-8b推理报错

[william-ljz]

很好的工作！
我在用timelens-8b推理的时候发现会报错如下， 想问一下作者的transformer版本或者改了啥吗，很奇怪我自己rl之后和官方的权重都能推理：

>> Traceback (most recent call last): |   -- | --line 959, in main |     |   | 2026-01-13 19:59:28.390656 master-0 >> dump_inputs = processor.apply_chat_template( |     |   | 2026-01-13 19:59:28.390657 master-0 >> File "/usr/local/lib/python3.10/dist-packages/transformers/utils/deprecation.py", line 172, in wrapped_func |     |   | 2026-01-13 19:59:28.39066 master-0 >> return func(*args, **kwargs) |     |   | 2026-01-13 19:59:28.390661 master-0 >> File "/usr/local/lib/python3.10/dist-packages/transformers/utils/deprecation.py", line 172, in wrapped_func |     |   | 2026-01-13 19:59:28.390663 master-0 >> return func(*args, **kwargs) |     |   | 2026-01-13 19:59:28.390674 master-0 >> File "/usr/local/lib/python3.10/dist-packages/transformers/processing_utils.py", line 1705, in apply_chat_template |     |   | 2026-01-13 19:59:28.390676 master-0 >> out = self( |     |   | 2026-01-13 19:59:28.390677 master-0 >> File "", line 170, in __call__ |     |   | 2026-01-13 19:59:28.390679 master-0 >> videos_inputs = self.video_processor(videos=videos, **output_kwargs["videos_kwargs"]) |     |   | 2026-01-13 19:59:28.390681 master-0 >> File "/usr/local/lib/python3.10/dist-packages/transformers/video_processing_utils.py", line 206, in __call__ |     |   | 2026-01-13 19:59:28.390683 master-0 >> return self.preprocess(videos, **kwargs) |     |   | 2026-01-13 19:59:28.390685 master-0 >> File "/usr/local/lib/python3.10/dist-packages/transformers/video_processing_utils.py", line 387, in preprocess |     |   | 2026-01-13 19:59:28.390686 master-0 >> preprocessed_videos = self._preprocess(videos=videos, **kwargs) |     |   | 2026-01-13 19:59:28.390688 master-0 >> File "/usr/local/lib/python3.10/dist-packages/transformers/models/qwen3_vl/video_processing_qwen3_vl.py", line 242, in _preprocess |     |   | 2026-01-13 19:59:28.39069 master-0 >> patches = patches.view( |     |   | 2026-01-13 19:59:28.390692 master-0 >> RuntimeError: shape '[1, 108, 2, 3, 22, 2, 16, 40, 2, 16]' is invalid for input of size 597196800

[jonasjunzhang]

Hi, we use transformers version 4.57.1 for TimeLens-8B, and we didn't do any further modification on the model compared to Qwen3-VL.

To avoid bugs and get the desired inference results, I strongly recommend you to follow our huggingface README to

1. install the packages (transformers, torch, etc)
2. perform inference with TimeLens-8B. The code on this page makes sure the fps and resolution aligns with training, so you can get the best results.

[zrli03]

我遇到了一次这个错误，当时用Qwen2.5VL的脚本来跑的，忘了改ele里的参数，改成跟这个仓库一样的：
parser.add_argument("--min_tokens", type=int, default=16) parser.add_argument("--total_tokens", type=int, default=3584)
之后就OK了