Xiangyu Zhao
Xiangyu Zhao
I find **refcoco-unc** in **tools/seqtr2coco/convert_mix_ref.py** :  but it's not in refcoco data structure: 
 category_id < num_classes, how to fix this?
> it's ok after changing the browser, thanks for your answer! The version of my Gradio is 4.16.0 and I have lots of trouble in using ```llava.serve.gradio_web_server```. Can you tell...
I'm stuck in building mmcv environment. Can you please tell me the version of your torch, torchvision, mmcv, mmdet and mmdet3d？I really appreciate it.
This problem hasn't been solved. This environment instruction is terrible !
你好，一般情况下我们默认推理Internvl-76B需要四张卡以**防止out of memory**。具体的切分方法在**split_model**函数中。
可以print `split_model()`函数的结果**device_map**和**visible_devices**看一下，是否能和本地的环境对齐
@dongyh20 Could you please assist in resolving this question?
你好，结果由于环境不同（transformers, torch, flash-attn, cuda）等会有不等的数值波动。 您的MMStar结果(59.6)和我们的结果(60.7)差距较小，可以理解为环境不同带来的小范围波动。 如果您有其他问题，欢迎讨论。
您好，对于您的Generation Config： ``` top_p=0.001, top_k=1, temperature=0.01, repetition_penalty=1.0, ``` 我们在评测的时候一般仅设置temperature=0，用于确保结果的稳定性。此外，torch版本和transformers的版本也会造成影响。