batch size matters in training

[WizardMx]

in adapt_llava.py, you have this

for j in range(8):
                        cur_image_features = image_features[cur_image_idx+j*4]

I don't quite understand this:
First, cur_image_idx keeps being 0, which means you don't support multi-image training case
Also, I think 4 there should be consistent with the batch size, not always 4.

Somebody has the same confusion with me?