Distributed task fails immediately
When a training task on zCLOUD, the GPU cloud, is set to PyTorch or MPI, check its framework: the framework sets what kind of job the task is.
Check the framework
A training task has a framework, and the framework sets the kind of job:
- **Simple** is a single-node job with one worker.
- **PyTorch** and **MPI** are for multi-node, distributed jobs.
On the default task form, the **Maximum Number of Nodes** field explains the link between node count and job type in its tooltip: "When the number of nodes is 1, it is a simple task, and when the number of nodes is greater than 1, it is a pytorch task."
If the job runs on a single node, use **Simple**. A job meant to be distributed uses PyTorch or MPI.
Change the framework
The framework is chosen when the task is created, on a step that the task form shows only once **Advanced Settings** is on.
- Open [**Training**](https://console.zettabyte.com/task-training/index), select **Create Task**, and turn on the **Advanced Settings** switch.
- On the **Instance Configuration** step, select the framework: **Simple**, **PyTorch** or **MPI**.
Next steps
- See [Execution frameworks in Run a training job](https://docs.zettabyte.com/docs/zcloud/tasks/run-a-training-job) for each framework and the rest of the task form.
- See [Common issues](https://docs.zettabyte.com/docs/zcloud/troubleshooting/common-issues) when the task still does not run after the framework is set.