GPU inference memory optimization.

Reduces GPU memory requirements for large model inference by orders of magnitude with no loss in output quality.