Google Cloud

Run SelfJev on a GPU-backed Compute Engine VM. This is a manual deployment path; the project has no GCP provisioning command and has not validated this guide end to end.

1. Provision the host

Request GPU quota in a region with capacity. A G2 VM with one NVIDIA L4 (24 GB VRAM) is a reasonable serving starting point. Check the selected machine’s system RAM, disk, and current hourly cost before creating it. L4 latency is not benchmarked for the current model.

Use a GPU-ready Linux image, or install a compatible NVIDIA driver using Google’s guide. Verify nvidia-smi on the host before installing SelfJev. Budget at least 50 GB free disk for dependencies and model files.

2. Install SelfJev

SSH to the VM and install Git, Git LFS, and uv, then follow the quickstart. Keep the service on loopback during initial testing:

export SELFJEV_API_KEYS="replace-with-a-long-random-key"
uv run --no-sync selfjev serve --host 127.0.0.1 --port 8000

This command runs from the prepared SelfJev checkout. The base model downloads on first start. For a persistent endpoint, supervise the process with systemd or use the Docker deployment with a restart policy.

3. Test through a tunnel

From your local machine, use your actual instance and zone:

gcloud compute ssh INSTANCE_NAME --zone ZONE \
  -- -N -L 8000:127.0.0.1:8000

Keep the tunnel open and point the SDK at http://localhost:8000. A production public endpoint needs HTTPS termination, authentication, and appropriate firewall rules. Do not expose the unauthenticated service to the internet.

4. Delete trial resources

Delete the VM when done, and review persistent disks, reserved addresses, and other resources separately. Stopped VMs may retain chargeable storage.

Provider references: accelerator-optimized machines, including G2, GPU driver installation, and gcloud compute ssh. This guide describes a deployment plan, not a measured GCP configuration.