Convert model weights to another data type.

• less model size on gpu global memory
• faster inference