Official implementation of GradES - a gradient-based selective training method that dynamically freezes converged modules during fine-tuning to achieve 40-50% computational savings without sacrificing ...