Increase efficiency for inference

[repoit]

Any ideas on how the inference speed for computing matches can be speed up? I already did some research (about parallelization, TensorRT, or reducing floating point precision) but am unsure of the best direction.