How did you manage to set the batch size to 16?

[six-finger]

Thanks for your outstanding contribution! I attempted to replicate your work using four NVIDIA RTX 2080Ti GPUs, with default parameters of crop_size=320480 and resized_size=256256. However, when I set the batch size to 1, the memory usage of one GPU already peaked at around 3000 / 11264 MB. This raises the question: how did you manage to set the batch size to 16?