FBLayout Optimizes Mobile GPU Memory for Transformer Finetuning

FBLayout introduces a unified R-Tile layout and tile-based index transformations to reduce memory overhead during transformer finetuning on mobile GPUs. The framework co-designs tensor organization with hardware to eliminate physical data movement during forward and backward passes.