ZeroGPU Spaces: Implement Ahead-of-Time Compilation for Faster Inference

AI Impact Summary

ZeroGPU Spaces are experiencing performance bottlenecks due to PyTorch’s just-in-time compilation, which isn’t optimized for the short-lived, frequently spun-up processes of the platform. Ahead-of-time (AoT) compilation using PyTorch’s `torch.export` and `torch.compile` offers a solution by allowing models to be optimized once and instantly reloaded, resulting in significantly faster demo generation times – up to 1.3x-1.8x speedups. This change introduces a more efficient workflow for deploying computationally intensive models like Flux, Wan, and LTX within ZeroGPU Spaces.