Yeah, tokens/s varies a lot based on workload. However, I’ve calibrated the estimator against public benchmarks, so it stays within a 30% error margin!

I'll definitely explore how to estimate vision models next!