Self-Hosted Python & AI/ML GitHub Actions Runners
Short answer: Hosted CI lacks unified Apple Silicon memory, private GPUs, and direct access to internal datasets. Strict turns local Macs and Linux GPU workstations into zero-account runners inside your private tailnet.
Run Python test suites, local MLX/PyTorch workflows, and private dataset checks on Macs and GPU Linux boxes.
Drop-in workflow YAML & zero-sudo enrollment
Enroll your host with strict add-runner and route jobs directly via GitHub's native runs-on labels:
jobs:
ml-eval:
runs-on: [self-hosted, Linux, X64]
steps:
- uses: actions/checkout@v4
- run: python3 -m unittest discover
curl -fsSL https://strict.ai/install.sh | sh
strict add-runner owner/ml-repo --labels self-hosted,Linux,X64,gpu
How Strict keeps Python & AI/ML runners healthy
| Subsystem | How it applies to Python & AI/ML |
|---|---|
| Disk Cleanup | Tiers 1–5 sweep runner temp directories, package stores, git-ignored scratch artifacts, and idle Docker containers whenever disk tightens. |
| Preflight Doctor | disk-headroom protects large model checkpoints from filling SSDs; pressure alerts trigger at >= 90% RAM or high 5m CPU load. |
| Tracked Toolchains | python, docker, node, rust, go, runner |
| Cache & Balancing | computeFleetDirective drains slots under high RAM pressure (>= 85% enter / <= 75% exit) so concurrent ML jobs don't OOM-kill each other. |
When Strict fits vs. when hosted runners win
Use Strict for Python & AI/ML when
- Direct access to Apple Silicon unified RAM (MLX) and on-prem Linux GPUs
- Model weights and proprietary evaluation datasets stay inside your Tailnet
- RAM hysteresis load balancing prevents OOM crashes on parallel test runs
Use cloud-hosted runners when
- Public fork PRs on open-source Python packages
- Occasional massive cloud GPU training bursts outside local capacity
Frequently asked questions
How does Strict prevent memory exhaustion during parallel Python or ML jobs?
computeFleetDirective monitors RAM usage continuously and starts draining secondary runner slots above 85% RAM (exiting below 75%) while firing pressure warnings at 90%.
Do proprietary weights or datasets leave my network?
No. Strict never proxies build traffic; jobs run directly on your own hosts inside your local LAN or Tailscale tailnet.
Can Strict manage both Apple Silicon MLX Macs and Linux GPU servers?
Yes. Both platforms run the same zero-sudo Python agent and report unified health, disk headroom, and toolchain telemetry.
Try Strict on one spare Mac in under two minutes
Signed Apple Silicon app or single-line CLI · Zero accounts required