Question about Training SAM 2.1 on Custom Image Dataset with RTX 3090

Dear SAM 2.1 Team,

I am writing to seek guidance on training the SAM 2.1 model on my custom image dataset. I am working in Visual Studio Code and have an NVIDIA RTX 3030 GPU with 24GB of RAM. My dataset consists of images and corresponding segmentation annotations.

I have been following the training documentation and have a few questions regarding the necessary modifications to the configuration files and training scripts to adapt the process for my specific setup.

Specifically, I would appreciate your insights on:

Which .yaml configuration file would be the most suitable starting point for training on a custom image dataset?
What are the essential changes I need to make in the chosen .yaml file to point to my image data (image folder, annotation folder, file list, etc.)?
Are there any specific parameters in the .yaml file (such as batch size, number of workers, etc.) that I should pay particular attention to given my RTX 3030 GPU with 6GB of VRAM? I understand this has less VRAM than the A100 mentioned in the documentation.
Do I need to make any modifications to the train.py script or other files for custom image dataset training?
I have already made some initial adjustments based on the documentation, such as setting the number of frames to 1 for image training and trying different batch sizes. However, I would greatly value your expert advice to ensure I am on the right track.

Thank you for your time and assistance.