Brian Qu

Hello, I think this is because that you didn't set the scope. You can try to change `PackInputs` to `mmpretrain.PackInputs`. It works for me.

I'm also curious about the concrete reason. I find that `configs/__base__/default_runtime.py` has a key named `default_scope` whose value is `mmpretrain`. I guess it's the reason taht `PackInputs` works when I...

It's been a long time. I remember that I've modify the token_id setting in `generate` function of InstructBLIP. Actually, this doesn't have much impact at all.

是在哪个setting跑的呢，然后loss趋于平缓也很正常，并不是说loss基本不动就会停止啊

I've pushed the loss curves in README. I'll close this issue.

感谢关注！issue标题里的`Qwen7B-chat/None`的`/None`是？然后对于您后续的问题，我感觉应该是cuda版本的问题 你可以试试 ``` import torch torch.cuda.is_available() ``` 版本库依赖方面可以参考[requirements](https://github.com/Coobiw/MiniGPT4Qwen/blob/master/requirements.txt) 您这个问题建议参考：[https://stackoverflow.com/questions/57814535/assertionerror-torch-not-compiled-with-cuda-enabled-in-spite-upgrading-to-cud](https://stackoverflow.com/questions/57814535/assertionerror-torch-not-compiled-with-cuda-enabled-in-spite-upgrading-to-cud)

看起来似乎是下载Qwen模型或者transformers版本的问题，方便的话可以给一个你的Qwen7B-Chat目录内容的截图吗

推理上应该不太存在bubble time的问题，这个主要是训练上的 现在推理没支持PP，但是支持了mapping到multi-gpu上（采用了transformers的`device_map="auto"`的方法实现） 可以参考刚才的commit：[https://github.com/Coobiw/MiniGPT4Qwen/commit/683326bcba26ca6dfd516f5ef48f2c35e71b2c0d](https://github.com/Coobiw/MiniGPT4Qwen/commit/683326bcba26ca6dfd516f5ef48f2c35e71b2c0d)

It's been a long time. I'll close this issue.

@echo840 感觉是不是因为这里的脚本是基于logits的，我看llava1.5的库里是直接基于generate出的文本的，mmbench官方的也是使用的后者，基于文本会考虑指令遵循的能力