GPU Server User Manual
1. The One Thing to Remember
Always launch GPU jobs using gpurun.
gpurun python train.py
2. GPU-hour Budget
- Weekly GPU-hour quota.
- Resets every Monday.
- Check using
gpurun --status.
3. Running Jobs
Python
conda activate myenv gpurun python train.py
Multiple GPUs
gpurun -g 2 torchrun --nproc_per_node=2 train.py
Docker
gpurun --docker -v $HOME/project:/work pytorch/pytorch python /work/train.py
Jupyter
gpurun jupyter lab --no-browser
4. Status
gpurun --status gpurun nvidia-smi
5. Queue Behaviour
Jobs wait automatically if GPUs are busy. Higher-priority users are served first.
6. Out of Budget
Preemptible jobs may run on idle GPUs and receive a 60-second grace period before termination.
7. Stop Jobs
gpurun --kill <job_id>
8. Quick Reference
| Task | Command |
|---|---|
| GPU Job | gpurun python train.py |
| 2 GPUs | gpurun -g 2 python train.py |
| Docker | gpurun --docker IMAGE CMD |
| Status | gpurun --status |
| Kill | gpurun --kill JOBID |
By using the GPU Server you agree to comply with this policy.