Currently, the Evaluator class for text-classification computes the model metrics on a given dataset. In addition to model metrics, it would be nice if the Evaluator could also report runtime metrics like eval_runtime (latency) and eval_samples_per_second (throughput).

cc @philschmid