Run Python Workloads on Hugging Face Jobs

This skill runs Python workloads on Hugging Face Jobs cloud infrastructure and manages the authentication tokens those jobs need. It points to a detailed reference guide whose safety, prerequisite, and validation requirements are mandatory, and lists a prerequisites checklist: a Hugging Face account on a Pro, Team, or Enterprise plan (Jobs require a paid plan), an authenticated login verifiable with hf_whoami(), and an HF_TOKEN for any Hub operation such as pushing models/datasets or downloading private repos. It distinguishes read, write, and organization token types, and states tokens are required for pushing to the Hub, accessing or creating private repositories, or programmatic Hub API use, but not for downloading public models/datasets or running jobs with no Hub interaction. There are three ways to provide a token to a job: the secrets={"HF_TOKEN": "$HF_TOKEN"} placeholder (recommended - auto-replaced with the logged-in session's token by the hf_jobs MCP tool, encrypted server-side), an explicit hardcoded token (not recommended, visible in code/logs), or an env variable (less secure than secrets, visible in job logs). It flags a specific pitfall: the $HF_TOKEN placeholder is only auto-replaced by the hf_jobs MCP tool - calling HfApi().run_uv_job() instead requires passing the real token via get_token(), since passing the literal string "$HF_TOKEN" produces a 9-character invalid token and a 401 error.

Use it to run Python workloads on cloud infrastructure without local GPU/TPU setup, process data at scale, run batch inference or experiments, schedule recurring tasks, use GPUs/TPUs for any workload, or persist results to the Hugging Face Hub. It does not apply outside Hugging Face Jobs workload and token management - a different domain or tool is out of scope.

Inputs: a Python script (as a uv script with inline dependencies), a compute flavor and timeout, and, when Hub access is needed, an HF_TOKEN passed via secrets. Outputs: a completed cloud job run, and, when the script pushes to the Hub, an uploaded model or dataset. It also documents fixes for common token errors: 401 Unauthorized (token missing or invalid - add the secrets config), 403 Forbidden (token lacks required permissions - use a write-scoped token), token not found in environment (wrong key name - use secrets not env), and repository access denied (token's account lacks access to the private repo).

The Hugging Face Hub and its huggingface_hub Python library (HfApi, whoami, get_token), the hf_jobs MCP tool and its secrets/env/flavor/timeout job parameters, and HfApi().run_uv_job() as an alternative to the MCP tool.

from huggingface_hub import whoami
user_info = whoami()  # Returns your username if authenticated

Engineers running Python workloads on Hugging Face's cloud GPU/TPU infrastructure who need to authenticate correctly and avoid the token-exposure and 401/403 pitfalls of hardcoded or environment-variable tokens.