Hugging Face ZeroGPU Builder
Build dependable Gradio AI demos and GPU workloads on Hugging Face ZeroGPU.
The skill explicitly addresses process isolation, pickle boundaries, CUDA-tensor returns, concurrent state, and fixed-path data leakage or overwrites, with a tempfile mitigation. It does not provide credential handling, least-privilege controls, user confirmation, rollback, or complete data-flow disclosure; it also relies on direct URLs and external Hub services, so points are deducted.
The guidance on loading, quotas, concurrency, dependencies, and failure types is fairly consistent and diagnostically useful. However, this is a static-only review with no skill-specific execution tests or reproduction evidence, limited abnormal-input coverage, and exposure to runtime-version drift; the score is conservatively capped below the static threshold.
The frontmatter gives precise triggers such as ZeroGPU, @spaces.GPU, import spaces, and named errors, and limits applicability to Gradio SDK Spaces while excluding Docker and Static Spaces. Chinese-language support is not evidenced, mainland-China reachability of core Hugging Face services is not established, and boundaries for non-Gradio or complex workloads remain incomplete, so points are deducted.
The skill has frontmatter, scope, progressive reference files, examples, anti-patterns, limitation disclosures, and dependency guidance. Repository context supplies Apache-2.0 licensing, CI validation, and a pinned revision, but there is no skill-specific version policy, changelog, named maintenance owner, or explicit update path, and many rules depend on changing external documentation.
It covers the main ZeroGPU implementation concerns—duration, quota, process boundaries, concurrency, CUDA builds, and caching—and includes directly usable patterns. Static review provides no execution or representative-output evidence, and some claims depend on current platform behavior, so only limited core-task usefulness can be credited under the static cap.
The files provide concrete rules, code snippets, error examples, reference paths, and repository CI checks for generated artifacts. The supplied CI does not demonstrate tests of ZeroGPU behavior, runtime mechanisms are not independently corroborated in the material, and no reproduction record is included, so evidence remains limited.
- This is a low-confidence static assessment; no code, dependency installation, or ZeroGPU behavior was executed or verified.
- ZeroGPU quotas, hardware, Python/runtime versions, and external documentation may change; verify current official documentation before use.
- The skill depends on Hugging Face Spaces, model downloads, and potentially external wheel or Hub URLs, requiring separate review of reachability, supply-chain risk, and sensitive-data boundaries.
- No evidence was provided for Chinese-language support or mainland-China network availability.
What does this skill do, and when should you use it?
This is one focused Agent Skill in the Hugging Face Skills monorepo, designed for Gradio SDK Spaces using ZeroGPU hardware. It guides developers through `@spaces.GPU`, Python and dependency configuration, process isolation, concurrency, quota behavior, CUDA build constraints, and model loading. It also covers large versus xlarge sizing, dynamic duration callables, `gr.State` semantics across workers, and example caching. It does not target Docker, Static, or Streamlit Spaces; the overall repository is licensed under Apache-2.0.
Guides the authoring and review of Python code containing import spaces or @spaces.GPU; explains module-scope model loading and CUDA device registration; helps configure duration, dynamic duration estimates, and size; documents Pickle boundaries, gr.State, concurrent requests, and unique output paths; explains python_version, requirements.txt, PyTorch version matching, and prebuilt CUDA wheels; and shows how to reproduce ZeroGPU example-caching behavior locally.
- A developer building a Gradio inference demo on Hugging Face ZeroGPU who needs the correct `@spaces.GPU` structure.
- A Space maintainer debugging `PicklingError`, `illegal duration`, `quota exceeded`, or CUDA wheel build failures.
- An application author estimating runtime from variable inputs to reduce quota pre-check failures.
- An engineer deploying models to ZeroGPU who needs to prevent shared-state corruption and output-file collisions under concurrency.
- A developer installing CUDA-dependent packages such as `flash-attn` when the build environment lacks `nvcc`.
What are this skill's strengths and limitations?
- Addresses ZeroGPU-specific quota, process-isolation, concurrency, and CUDA build problems.
- Provides actionable Python, README frontmatter, and dependency-configuration patterns.
- Clearly distinguishes ZeroGPU from dedicated GPU, CPU, Docker, and Static Space environments.
- Calls out easily missed constraints such as minimum realistic duration, eager model loading, and CPU-safe return values.
- Its scope is limited to Gradio SDK Spaces using ZeroGPU, rather than general Gradio or general Hugging Face deployment.
- The backing GPU, supported Python versions, and tier limits can change, so current ZeroGPU documentation remains necessary.
- Detailed concurrency, quota, lifecycle, and CUDA dependency guidance relies on the repository’s referenced files.
- The supplied source does not provide a test suite or platform compatibility matrix, so target-Space validation is still needed.
How do you install this skill?
Copy or symlink skills/huggingface-zerogpu from the repository into a standard Codex .agents/skills location, such as $REPO_ROOT/.agents/skills or $HOME/.agents/skills. The README also documents installation through the Hugging Face CLI with hf skills add <skill-name>; replace the placeholder with the target skill name.
How do you use this skill?
In a coding agent that supports the skill, ask for a concrete task such as: “Review this @spaces.GPU Gradio Space for duration settings, Pickle-safe returns, concurrency safety, and CUDA dependencies.” You can also provide code containing import spaces or @spaces.GPU; those are explicit activation signals in the skill description.
How does this skill compare with similar options?
Compared with the repository’s huggingface-gradio skill, this skill focuses on ZeroGPU hardware, quota, process, and CUDA constraints rather than general Gradio components and event listeners. Compared with dedicated-GPU Spaces, it describes request scheduling, GPU slicing, and quota behavior; Docker and Static Spaces are outside its scope.