Run your GitHub Actions workflows on Hugging Face Jobs with a one-line change:
jobs:
train:
runs-on: hf-jobs-a10g-small # was: ubuntu-latest
steps:
- uses: actions/checkout@v4
- run: python train.pyGPU CI for the price of a Hub subscription. No self-hosted runner infrastructure to babysit.
Note that Jobs-Actions is in Beta, and could change. But the overall motivation for this is that GitHub-hosted runners are becoming increasingly unreliable. HF Jobs instances are reliable and offer GPU machines at much more reasonable, per-minute prices. This bridges the GitHub and HF Jobs: GitHub Actions stays in charge of orchestration, YAML, secrets, and the Checks UI; HF Jobs provides the compute.
| GitHub-hosted | jobs-actions (HF Jobs) | Self-hosted on EC2 | |
|---|---|---|---|
| GPU access | ❌ (limited beta only) | ✅ T4, A10G, A100, H200, L4, L40s | ✅ DIY |
| Cold start | ~5–15s | ~30–90s | ~0s (warm pool) |
| Idle cost | $0 | $0 (ephemeral) | $$ |
| Setup | None | 5-minute one-time install | Days |
PR push
→ GitHub fires `workflow_job.queued` webhook
→ Dispatcher (HF Space) receives it, sees `runs-on: hf-jobs-*` label
→ Mints a one-shot GitHub Actions runner registration token
→ Launches an HF Job with a base image (or your prebuilt runner image)
→ Runner registers ephemerally, GitHub dispatches the job to it
→ Runner exits after one job, container terminates
| Label | HF Jobs flavor | Notes |
|---|---|---|
hf-jobs-cpu |
cpu-basic |
2 vCPU, 16GB RAM |
hf-jobs-cpu-upgrade |
cpu-upgrade |
8 vCPU, 32GB RAM |
hf-jobs-cpu-performance |
cpu-performance |
Performance CPU tier |
hf-jobs-cpu-xl |
cpu-xl |
Highest CPU tier |
hf-jobs-t4-small |
t4-small |
T4 GPU (cheapest) |
hf-jobs-t4-medium |
t4-medium |
T4 GPU, more RAM |
hf-jobs-l4x1 |
l4x1 |
1× L4 GPU |
hf-jobs-l4x4 |
l4x4 |
4× L4 GPU |
hf-jobs-a10g-small |
a10g-small |
A10G GPU |
hf-jobs-a10g-large |
a10g-large |
A10G, more RAM |
hf-jobs-a10g-largex2 |
a10g-largex2 |
2× A10G |
hf-jobs-a10g-largex4 |
a10g-largex4 |
4× A10G |
hf-jobs-a100-large |
a100-large |
A100 GPU |
hf-jobs-l40sx1 |
l40sx1 |
1× L40s |
hf-jobs-h200 |
h200 |
H200 GPU |
Apache 2.0