Skip to content
Kitbay

huggingface-zerogpu

Apache-2.0Updated on Nov 11, 202611,334GitHub stars

A Hugging Face skill for writing and reviewing Gradio Spaces that run on ZeroGPU hardware.

What the skill teaches

A Hugging Face skill for writing and reviewing Gradio Spaces that run on ZeroGPU hardware. It explains the @spaces.GPU decorator, the two GPU sizes, how CUDA is available only inside decorated calls, choosing call durations against quota, pickle-based process isolation between the main process and GPU workers, state handling across that boundary, concurrency, why torch.compile is replaced by ahead-of-time compilation, CUDA build constraints and dependency setup. It applies only to Gradio SDK Spaces, and reference files go deeper on concurrency, quota, how ZeroGPU works, and CUDA dependencies.

Capabilities

  • Write Gradio Space code for ZeroGPU with @spaces.GPU
  • Tune GPU call duration against quota
  • Handle process isolation and state across GPU workers
  • Install CUDA-dependent packages for ZeroGPU