What problem does it solve? ZeroGPU Spaces behave differently from ordinary GPU environments: the GPU is only attached inside decorated functions, handlers run concurrently by default, quota is pre-checked against declared duration, and CUDA builds lack nvcc. This Skill prevents the subtle failures—quota rejections, pickle errors, hangs, and build breakage—that come from treating ZeroGPU like a normal GPU. ## Core Features & Use Cases - Correct @spaces.GPU usage: Covers module-scope model loading, duration and size tuning, dynamic duration callables, and why torch.compile must be replaced with AoTI. - Isolation and concurrency safety: Explains pickle boundaries, gr.State copy semantics, the ban on returning CUDA tensors, and rules against mutable globals and fixed file paths. - Dependency and build constraints: Details wheel-only CUDA installs, wheel filename tag reading, torch side-car pinning, and the python_version frontmatter requirement. - Use Case: You are building a Gradio image-generation demo on ZeroGPU and users hit "quota exceeded" after a few runs. This Skill shows you to declare a realistic duration (e.g. 15s instead of the 60s default) so requests stop being rejected prematurely. ## Quick Start Ask the assistant to review my ZeroGPU Gradio app for correct @spaces.GPU usage, duration settings, and concurrency safety.
