Skip to main content

Datacenter & Infrastructure Issues

GPU & Driver Issues​

nvidia-smi returns "No devices were found"​

After deploying a VM, the GPU hardware is not detected.

Solutions:

  1. Reboot the instance — This resolves most transient GPU detection issues.
  2. If the issue persists after reboot, the GPU may have a hardware problem on the host. Try renting from a different provider or datacenter.

"Failed to initialize NVML: Driver/library version mismatch"​

Ubuntu may auto-update GPU drivers, causing a mismatch between the loaded kernel module and the userspace library.

Solution: Reboot the instance to load the matching driver version:

sudo reboot

CUDA initialization errors in PyTorch​

If PyTorch shows "CUDA initialization: CUDA unknown error" or cannot detect the GPU:

  1. Verify the GPU is visible: nvidia-smi
  2. If nvidia-smi works but PyTorch doesn't detect the GPU, check your CUDA toolkit version matches the driver version.
  3. Reboot the instance and try again.
  4. If the issue persists, try a different instance — the GPU on the current host may have a hardware issue.

High VRAM usage on a fresh container with no processes​

In container mode, the GPU is shared between users on the same node. Other users' processes may be consuming VRAM.

Solution: Use VM mode instead of container mode for dedicated GPU access. In VM mode, the GPU is passed through exclusively to your instance.

Networking Issues​

Cannot access services between VMs​

If a service running on one VM (e.g., port 8080) is not reachable from another VM:

  1. Check binding address — Make sure the service binds to 0.0.0.0, not 127.0.0.1 or localhost.
  2. Same network — VMs on different subnets may not be able to communicate directly. Contact support to ensure your VMs are deployed on the same network.

Cannot access a web service from your browser​

If you start a web service (e.g., ComfyUI, Jupyter) but can't reach it from your local browser:

  1. Bind to the public IP — Many services default to localhost. Launch with the --listen or --host flag:
    # Example for ComfyUI
    python main.py --listen 0.0.0.0 --port 8188
  2. Access the service at http://<vm_public_ip>:<port>.

Port availability​

All ports are open by default on VMs — there are no firewall restrictions from CloudRift's side. You can configure your own firewall rules using ufw or iptables on the VM.

Docker Container Mode Issues​

Wrong container spins up​

If selecting "No Container" still launches a pre-configured container:

  1. Log out and log back into the CloudRift Console.
  2. Retry the deployment.

Cannot connect to container IP​

Container mode is less mature than VM mode. If you cannot connect to a container instance:

  1. Verify the instance status is "Ready" in the console.
  2. Try VM mode instead, which provides a full Linux environment with SSH access.

System Service Issues​

Service not starting​

Check the service logs:

sudo journalctl -u rift
sudo systemctl status rift

Docker permission errors​

Ensure the rift service is running as root and Docker is accessible:

sudo systemctl status docker
docker info

If Docker is not running:

sudo systemctl start docker
sudo systemctl enable docker

GPU not detected by the service​

  1. Verify the GPU is visible to the host: nvidia-smi
  2. Ensure you rebooted after installing or removing drivers.
  3. Check VFIO binding if running in VM mode: lspci -k | grep -A 2 -i nvidia