微调internXC2报错

[yuzeng0-0]

File "/home/ma-user/anaconda3/envs/dpo/lib/python3.10/site-packages/transformers/configuration_utils.py", line 264, in getattribute return super().getattribute(key) AttributeError: 'InternLMXcomposer2Config' object has no attribute 'image_token_index'

[TideDra]

我们的实现方法加入了单独的image token标记图片位置，这在原模型中是没有的。您需要在模型的config.py文件中添加"image_token_index":92544

[yuzeng0-0]

感谢

[yuzeng0-0]

File "/cache/zy/VL-RLHF/src/vlrlhf/models/InternLMXC2/init.py", line 87, in _merge_input_ids_with_image_features raise ValueError( ValueError: The input provided to the model are wrong. The number of image tokens is 2 while the number of image given to the model is 2. This prevents correct indexing and breaks batch generation.这个怎么解决

[yuzeng0-0]

File "/cache/zy/VL-RLHF/src/vlrlhf/models/InternLMXC2/init.py", line 87, in _merge_input_ids_with_image_features raise ValueError( ValueError: The input provided to the model are wrong. The number of image tokens is 2 while the number of image given to the model is 2. This prevents correct indexing and breaks batch generation.这个怎么解决 if image_to_overwrite.sum() != image_features.shape[:-1].numel():条件判断raise ValueError

[TideDra]

我们统一用<image>作为image token，请检查下输入文本中image token的总数量与输入图片的总数量是否一致

[yuzeng0-0]

我们统一用<image>作为image token，请检查下输入文本中image token的总数量与输入图片的总数量是否一致这个是没问题的。 我猜测是padding的问题，internXC2的pad_token_id是2，对应embedding为零向量，在执行final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)的时候报错

[yuzeng0-0]

方便展示一下internXC2模型的config.py文件吗

[TideDra]

方便展示一下internXC2模型的config.py文件吗

可以参考我们放出的checkpoint: https://huggingface.co/TideDra/internlm-xcomposer2-vl-7b-DPO/blob/main/config.json

[yuzeng0-0]

internXC2 的DPO代码还是跑不起来，您确定repo里是最新版本的代码吗？是不是repo是之前的版本导致有些bug没有修复