TRAINING VGGT fails with CUDA OOM

[OumaimaCharia]

I’m trying to retrain VGGT weight with a small dataset (31 PNG images, 480×640) on an RTX 2000 Ada (8 GB) and also on Colab (basic plan). Even with batch size = 1, reduced workers, and several frozen layers, training starts but fails with CUDA out of memory during validation (on Colab it happens even before training begins).

[jytime]

Hey 8GB would be too small for training.