How to add image data augmentation

[wufan-tb]

I am seeking assistance regarding an issue I'm encountering.

I intend to incorporate image data augmentation into the SFT process within ms-swift. Currently, this augmentation logic is integrated within the Template._encode function.

However, when training with a substantial dataset, while the mapping and packing stages proceed without incident, the training phase consistently results in a CUDA OOM error. I have confirmed that this error does not manifest when using the same volume of data without augmentation, or when training with a smaller dataset.

Could you please elucidate the potential reasons for this behavior, and what modifications or strategies would you recommend to resolve this issue?

[Jintao-Huang]

Check if it's caused by images being too large.