Comment by S0y

How much does this speedup inference for the end user in terms of tk/s ?