hugging-face-jobs

You can run Python workloads on Hugging Face Jobs by submitting a script with a secure HF_TOKEN secret, enabling automatic hardware selection across GPU and CPU environments for data processing and inference.

Token management for Hugging Face Hub interactions works by differentiating secrets from environment variables, using HF_TOKEN secrets to authenticate safe job execution and persist results to prevent ephemeral data loss.

Yes, you can use vllm for model inference on Hugging Face Jobs. The environment supports dependencies like vllm, torch, and transformers to execute GPU-enabled batch tasks and experiments securely.

The best way to persist results from Hugging Face Jobs to the Hub is by using secure token-authenticated workflows that save datasets and models as outputs, preventing ephemeral job data loss after execution.

Yes, you need a valid SKILL.md frontmatter containing name and description fields to run Hugging Face Jobs workloads, while optional resources can be included in scripts, references, and assets directories.

Hugging Face Jobs results disappear after execution because job data is ephemeral. You must persist results to the Hugging Face Hub using secure token authentication to save datasets and models as outputs.