[BUG] Pre-training unstable

[orrzohar]

是否已有关于该错误的issue或讨论？ | Is there an existing issue / discussion for this?

• [x] 我已经搜索过已有的issues和讨论 | I have searched the existing issues / discussions

该问题是否在FAQ中有解答？ | Is there an existing answer for this in FAQ?

• [X] 我已经搜索过FAQ | I have searched FAQ

当前行为 | Current Behavior

I am trying to pre-train MiniCPM-V-LLaMA3. I tried the following:

• lr 1e-3 to 1e-5
• batch size 128 to 1024 However, I keep getting NaNs in/after the perceiver resampler.

期望行为 | Expected Behavior

Stable training

复现方法 | Steps To Reproduce

1. Clone this repo
2. download image pre-training datasets
3. do pretraining

运行环境 | Environment

- OS:
- Python:
- Transformers:
- PyTorch:
- CUDA (`python -c 'import torch; print(torch.version.cuda)'`):

备注 | Anything else?

No response