mplug-owl3 doing full model sft bug report #2158
Comments
Thank you for your reply! |
|
Are you using the model from Modelscope or from Hugging Face? |
|
我的基础模型是从hugginging face上面下载的。
…---Original---
From: ***@***.***>
Date: Sun, Sep 29, 2024 13:12 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [modelscope/ms-swift] mplug-owl3 doing full model sft bug report(Issue #2158)
使用的是modelscope的模型还是huggingface的模型
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: ***@***.***>
|
|
有可能是我用的是huggingface模型，但是我没有设置hf=1？
我下午再试一下，主要是Lora是正常的
…---Original---
From: ***@***.***>
Date: Sun, Sep 29, 2024 13:12 PM
To: ***@***.***>;
Cc: ***@***.******@***.***>;
Subject: Re: [modelscope/ms-swift] mplug-owl3 doing full model sft bug report(Issue #2158)
使用的是modelscope的模型还是huggingface的模型
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID: ***@***.***>
|
|
For the first issue, could you please pull the latest repository from Hugging Face and try again? I have added a function named |
When I update the image_processing_mplugowl3.py from the huggingface latest repo and set USE_HF=1, all 2 Problems are solved! Thank you so much for your quick and valuable help! |
Describe the bug
There are 2 bugs! But that seems like these bugs only occur when doing full parameter finetune. When I use Lora, that seems like no such bugs.
1. If I train 1 epoch and do evaluation in training time after one val process finished, the coming training procedure will failed. The error message is follows:
2. If I train the model from checkpoint, which means I use the args:
--resume_from_checkpoint output/mplug-owl3-7b-chat/v34-20240929-110829/checkpoint-2 \The error will occur as follows.The text was updated successfully, but these errors were encountered: