Cast transformer to fp16

[znadrich-qf]

If I am using en_core_web_trf for inference is there a way to cast the underlying transformer model to fp16 for better model throughput?