Migrating Your GitHub CI to Hugging Face Jobs

Key Takeaways

→ GitHub Actions keeps control of CI while actual jobs run on Hugging Face serverless infrastructure

→ Trackio CPU CI improved from about 1m40s to 1m10s, roughly 30 percent faster

→ A new GPU test suite passed in 45 seconds on an HF Jobs t4-small instance for under a cent

→ Switching is mostly a one-line change, from runs-on ubuntu-latest to hf-jobs-cpu-upgrade

→ Setup requires duplicating a dispatcher Space and installing a GitHub App that mints short-lived runner tokens

Hugging Face published a hands-on guide on June 9, 2026 showing how to keep GitHub Actions in charge of CI while running the actual jobs on Hugging Face Jobs. The writeup is based on a real migration of Trackio, Gradio’s experiment-tracking library.

What Hugging Face Jobs Adds

A Job is a command, a Docker image, a hardware flavor and optional secrets bundled together. That makes it a natural fit for CI, where steps are already command-driven and often benefit from picking exactly the right accelerator.

The Bridge Architecture

The huggingface/jobs-actions project turns a GitHub Actions job into an ephemeral self-hosted runner. A triggered workflow sends a signed webhook to a dispatcher Space, which verifies the event, mints a short-lived runner registration token and launches an HF Job on matching hardware.

Setup in Five Steps

Duplicate the dispatcher Space under your namespace using cpu-upgrade so webhooks are not missed
Create and install the GitHub App from the Space and store an HF_TOKEN secret
Set final dispatcher variables, including HF_NAMESPACE if jobs should bill elsewhere
Change runs-on from ubuntu-latest to hf-jobs-cpu-upgrade or a GPU label
Test with a minimal workflow and watch logs stream back to GitHub

The Results

For Trackio, the numbers were concrete: a GitHub ubuntu-latest baseline of about 1m40s, an HF Jobs CPU run on the Microsoft Playwright image at 1m10s (about 30 percent faster), and a GPU check on t4-small finishing in 45s for less than a cent.

Images, Logs and Next Steps

The guide notes that image choice matters. A bare ubuntu:22.04 was slower than GitHub’s loaded runner, while mcr.microsoft.com/playwright:v1.60.0-jammy and nvidia/cuda:12.4.0-runtime-ubuntu22.04 worked well. Logs can be pulled with hf jobs logs, and Jobs also supports volume mounts for datasets and models.

Frequently Asked Questions

What is Hugging Face Jobs?

Hugging Face Jobs runs a command or script on Hugging Face serverless infrastructure with a chosen Docker image and hardware flavor such as CPU or a t4-small GPU, with optional environment variables and secrets.

How do you connect GitHub Actions to Hugging Face Jobs?

You duplicate the jobs-actions-dispatcher Space, create and install a GitHub App from it, store an HF token as a Space secret, then change the workflow runs-on label to something like hf-jobs-cpu-upgrade.

Does this replace GitHub Actions?

No. GitHub Actions still orchestrates workflows and scheduling. It simply hands the actual job execution to a self-hosted runner that lives inside a Hugging Face Job.

Can it run GPU CI?

Yes. Using a GPU label such as hf-jobs-t4-small launches the job on real GPU hardware, which is difficult and expensive with default GitHub-hosted runners.