Xingjun.Wang

Please update your env settings: modelscope==1.14.0 datasets==2.18.0 or update to the latest docker image version.

感谢反馈！ 目前发现1.13.2和1.13.3均会有此现象，可先回退到1.13.1，我们将在下个小版本修复这个问题。

> `site-packages/modelscope/msdatasets/utils/hf_datasets_util.py` 这里面的的导入都失效很多了，没有更新吗 Hi，预计下周初会发布新版本，fix掉该问题。

This issue has been fixed and merged to master branch. Version 1.14.0 will be released soon.

# 1. Construct the dataset ```text train.jsonl (each line): {"query_id": "111", "query": "吃饭的猫猫1", "image_id": "222", "image": "/path/to/cat_1.jpg"} validation.jsonl (each line): {"query_id": "333", "query": "吃饭的猫猫2", "image_id": "444", "image": "/path/to/cat_2.jpg"} ``` #...

same problem

The training process needs 18GB(or above) GPU memory, we recommend you to use a GPU with 24GB memory like Nvidia A10, A100, V100, etc. Plus, we are developing train-free models,...

check available cuda: torch.cuda.is_available()

模型预计在lora training期间最高消耗约19G左右的GPU显存，所以推荐使用24GB 或以上的显卡。另外，我们最近在开展train free相关技术的研究，可以关注一下后续几个迭代。

尝试升级modelscope到1.14.0