Huaxin Zhang

For inference with HolmesVAD-7B, a 24GB 4090 GPU is required. If your GPU has less than 24GB of VRAM (e.g., 12GB), you can try multi-GPU inference or quantized inference (set...

你好，这是由于你的服务器无法连接huggingface导致的，你可以设置你服务器的网络确保可以正常访问huggingface。 如果无法设置网络，也可以按照如下步骤进行： - 1.确认HolmesVAD-7B已下载至本地 - 2.下载LanguageBind/LanguageBind_Image和LanguageBind/LanguageBind_Video_merge至本地，并将HolmesVAD-7B文件夹下中config参数中的`mm_image_tower`和`mm_video_tower`参数分别改为本地路径 以下是一个可供参考的服务器下载模型的python脚本-`download.py`： ```python import huggingface_hub from huggingface_hub import snapshot_download import os huggingface_hub.login("your-huggingface-access-token") models = ["LanguageBind/LanguageBind_Image", "LanguageBind/LanguageBind_Video_merge"] for model in models: local_dir = './checkpoints/' + model...

检查你使用的gradio版本，我们目前使用的是3.37.0 ```bash pip install gradio==3.37.0 ```

I met the same question when I run finetune_lora.sh, the loss suddenly increases during training.The only modification I made was to use half of the llava-v1_5_mix665k samples.