Larger batch size prevents out of memory error.

I have been experimenting with the fine-tuning of SAM3 model. I assumed since this model is very heavy, we ought to use the smallest possible batch size (which is 1), that resulted in OOM error.

Later I tried batch_size = 16, and OOM is gone!

Why is that?
Are there any sort of logic behind the scene?

Start by reproducing SAM3 fine-tuning with batch sizes 1 and 16, then investigate how the training path handles memory and batching. Done means providing a clear explanation for the differing out-of-memory behavior, ideally with any relevant usage guidance documented.

Written by the indexing model from the issue text.