Replicate: torch.compile caching improves Flux model startup times

AI Impact Summary

Replicate has significantly improved model startup times for Flux models using caching of `torch.compile` artifacts. This automatic optimization provides a 2-3x speedup, alongside the already existing 30%+ inference speed improvements. This enhancement reduces the operational overhead and improves the developer experience for Flux model deployments on Replicate.