Configure GPU slicing for EAS services - Platform For AI

GPU slicing partitions a physical GPU's computing power and memory across multiple EAS service instances, improving utilization and reducing deployment costs.

Prerequisites

GPU slicing requires all of the following conditions:

Resource type: EAS resource group or Lingjun resource quota.
Instance status: GPU instances in the resource group must be running (not starting or stopped).

Note
First-time GPU provisioning takes 8–10 minutes. Wait until the instance is ready before proceeding.

Configuration

Configure GPU slicing when you create or update a service.

Console

Log on to the PAI console. Select a region on the top of the page. Then, select the desired workspace and click Elastic Algorithm Service (EAS).
Create or update a service to open its configuration page.

In the Resource Information section, configure the following parameters. Other parameters are documented in custom deployment.

Parameter	Description
Resource Type	Select EAS Resource Group or Resource Quota.
GPU Slicing	Enable GPU slicing. Note If this option is unavailable, check Why is the "GPU slicing" option missing from the console?.
Deployment Resources	Single-GPU Memory (GB): Required. GPU memory allocated to each instance. Must be an integer. Used for instance scheduling. Important For `ml` resource specs, the memory unit is GB. For `ecs` resource specs, the unit is GiB. Computing Power per GPU (%): Optional. Percentage of computing power per instance. Integer from 1 to 100. Used for instance scheduling. Both settings apply together. For example, setting memory to 48 GB and computing power to 10% limits each instance to 48 GB of GPU memory and 10% of computing power.

After you configure the parameters, click Deploy or Update.

CLI

GPU slicing fields in a JSON configuration file:
```
{
    "metadata": {
        "gpu_core_percentage": 5,
        "gpu_memory": 20
    }
}
```
- gpu_memory: Maps to Single-GPU Memory (GB) in the console.
- gpu_core_percentage: Maps to Computing Power per GPU (%) in the console. Requires gpu_memory; otherwise ignored.
Important
For memory-based scheduling, omit the gpu field or set it to 0. Setting gpu to 1 allocates the entire GPU exclusively, and the gpu_memory and gpu_core_percentage fields are ignored.
Use the create or modify command from the command reference to deploy or update a service.

FAQ

Q: Why is the GPU slicing option missing?

Troubleshoot as follows:

Verify that Resource Type is set to EAS Resource Group or Lingjun Resource Quota.
Verify the resource group has GPU resources (GPU column value > 0).
Confirm the GPU instance is running and not in other states such as starting or stopped. If initializing, wait until ready.