infrastructurefactbullishQLoRA achieves strong accuracy on vision models while meeting on-device VRAM budgets of around 2GBComputer Vision27 Jul 2026http://arxiv.org/abs/2607.02158v1