Usage workflow
1
Select CES Instance Type
Select a GPU Instance or CPU Instance based on the task type.
- GPU Instances are suitable for training, fine-tuning, and inference tasks that require GPU acceleration.
- CPU Instances are suitable for data preprocessing, scheduling services, development and debugging, or lightweight services.
2
Prepare a Logon Credential
When creating a CES instance, you need to select a logon credential. When using Custom Password, you can set the password during instance creation. When using Key Pair, prepare Key Pairs first.If you need to access the instance through SSH or other methods, determine in advance whether you need to configure a public network access entry point later.
3
Prepare data and storage
Select a volume based on the task’s data access method.
- When multiple CES instances need to share datasets, model weights, or training outputs, you can use GPFS file storage and mount the file system to the CES instances.
- When you need to transfer datasets, model files, inference inputs, or training outputs between your local environment and CES instances, you can use OSS object storage and upload or download them using ossctl.
4
Prepare the Network Environment
Before creating a CES instance, confirm the network environment required by the instance. If you need to place the instance in a specified network, create a VPC in advance. For details, see Manage VPCs.If the task runs only in an internal network environment, you generally do not need to configure a public network access endpoint in advance.
5
Create a CES Instance
Log in to the GPU Cloud console, go to the Instance page, and create a GPU Instance or CPU Instance as required by the task.When creating an instance, you typically need to confirm the region, billing mode, instance specification, image, and Logon Credential. GPU Instance and CPU Instance offer different resource options. For configuration details, see Create GPU Instance or Create CPU Instance.
6
Configure Public Network Access
If you need to access an instance from the public network or provide services externally, you can create an Elastic IP and NAT rules based on the access method after the CES instance is created. Before configuring public network access, confirm the private IP address, port, and protocol of the instance to be exposed. For details, see Manage Elastic IP and Manage NAT Rules.
7
Connect to an Instance and Run Tasks
After the CES instance is created, use a logon credential supported by the instance to connect to it. After connecting, you can prepare the runtime environment, upload or mount data, and start training, inference, development and debugging, or batch processing tasks.If you use a GPU Instance, we recommend confirming that the GPU resources and driver environment are available before starting the formal task.
8
View Status
While a task is running, you can use monitoring to view the status of resources such as CPU, memory, and GPU, and use logs to view task output or application errors. For details, see View GPU Instance Monitoring, View CPU Instance Monitoring, and Query and Analyze Logs.If you need to continuously monitor resource status or exceptional events, you can configure a notifier. For details, see Create Custom Alert Rules.
9
Clean Up Resources
After the task is complete, decide whether to stop, renew, or delete the CES instance based on the billing method and data retention requirements. Before deleting the CES instance, ensure that any data to be retained has been saved to the GPFS file system, OSS, or another persistent location.