If you require high computing efficiency for tasks such as data processing or offline model inference in MaxCompute, you can purchase and use MaxCompute AI compute resources (GU Quota). These resources are designed to meet the demand for GPU computing power. This topic describes how to purchase and use these compute resources.
Usage notes
These resources are available for purchase in the China (Beijing), China (Hangzhou), China (Shanghai), China (Shenzhen), and China (Ulanqab) regions.
GPU AI compute resources are supported only in MaxFrame.
You can change only the number of GUs in an AI compute resource quota group. You cannot change the GU specification. You cannot split a level-1 AI compute resource quota into multiple level-2 quotas.
Procedure
Log on to the MaxCompute console, and select a region in the upper-left corner.
In the navigation pane on the left, choose .
On the Quotas page, click New Quota.
On the purchase page, set Product Type to AI Computing Resources.
On the MaxCompute-AI Compute Resources (Subscription) page, configure the following parameters:
Specification: The specification of the GU quota. For example:
gu.a10.c30g186.Compute Unit: Indicates the number of GUs in the GU Quota.
The adjustable step size for compute units varies based on the quota specification. The step size can be 1, 2, 4, or 8 GUs. The limits specified on the purchase page apply.
Click Buy Now, and follow the on-screen instructions to agree to the Terms of Service and complete the purchase. (The GU resource will be provisioned in about 10 minutes).
After the purchase is complete, return to the console. A quota group of the GPU AI compute resource type appears on the Quota Management page.
Usage guide
MaxCompute provides dedicated AI compute resources (GU Quota) to efficiently run AI and machine learning workloads. You can allocate these resources to jobs on demand using the level-2 quota mechanism. The following sections describe the key steps for using AI compute resources.
Check the quota status
Similar to MaxCompute general computing resources (CU Quota), you can specify AI compute resources (GU Quota) at the job level. A new GPU resource group must be initialized before you can use it. You can check the status of a level-2 quota on the Quota Management page in the console:
Running: The environment is ready and you can submit jobs.
Creating: The resource configuration is in progress. Wait for the configuration to complete before you use the resource.
Confirm the quota status before you submit a job to ensure that the job can be scheduled.
Use AI compute resources in a MaxFrame job
You can use the @with_running_options decorator to declare resource requirements. This lets you explicitly specify GU resources for an AI function or user-defined function (UDF) in MaxFrame. The following example shows how to do this:
@with_running_options(gu=1, gu_quota="your_gu_quotaname")
def your_ai_function():
"""
gu=1: Requests 1 GU (one GPU card).
gu_quota: Specify the name of an activated GU Quota that is in the "Running" state.
This function uses the @with_running_options decorator to specify the use of
1 GU (GPU Unit) of AI compute resources at runtime, requested from the quota group named 'your_gu_quotaname'.
"""
# The function body goes here. 'pass' is used as a placeholder.
pass
This configuration ensures that the function runs on a cluster with GPU capabilities. This approach is suitable for multi-modal data processing, model inference, and other GPU-accelerated scenarios.