VRAM

The dedicated memory on a GPU that holds model weights, activations, and data during training or inference. Running out of it is the single most common reason a model training/inference job crashes.