You can deploy applications that require GPU-accelerated instances as functions using container images. This approach is ideal for popular AI projects, such as Stable Diffusion WebUI, ComfyUI, retrieval-augmented generation (RAG), and TensorRT. Using container images to deliver functions improves development and delivery efficiency.
Create a function
Log on to the Function Compute console. In the navigation pane on the left, choose .
In the top menu bar, select a region. On the Function List page, click Create Function.
In the dialog box that appears, select GPU Function and then click Create GPU Function.
On the Create GPU Function page, set the following parameters and then click Create.
Basic Configurations: Enter a Function Name. The name must be unique within the same Alibaba Cloud account and region, and must follow the naming conventions.
Elastic Configurations: Select an instance type. You cannot use provisioned instances and on-demand instances at the same time. After the function is created, you cannot change the instance type.
On-demand instances
Configuration Item
Description
Example
Instance Type
Select On-demand Instance. Instances scale automatically based on request volume and are released when there are no requests. You are billed for what you use.
On-demand Instance
GPU Card Type
Select a GPU card type. For more information about the specifications supported by different card types, see Instance types and specifications.
Ada series
Specifications
Set the GPU Memory, vCPU, Memory, and Disk specifications for the function based on your business needs. After you set the specifications, the usage of each resource is calculated by multiplying the specification by the duration of use. For more information, see Billing overview.
NoteAll directories on the disk are writable. The disk space is shared.
The disk is tied to the instance lifetime of the underlying function. When the system reclaims the instance, data on the disk is lost. If you need persistent storage, you can mount a NAS file system or an OSS bucket. For more information, see Configure a NAS file system and Configure Object Storage Service.
GPU Memory: 48 GB
vCPU: 8 vCPU
Memory: 64 GB
Disk: 512 MB (not billed, Function Compute provides a free quota of 10 GB disk space)
Minimum Instances
If your business is latency-sensitive, after you select Elastic Instance, we recommend that you set the minimum number of instances to 1 or greater to lock resources in advance and reduce cold start latency.
NoteAfter you set Minimum Instances to 1 or more, if no elastic policy for the minimum number of instances is configured or if no elastic policy is active for a period, the current minimum number of instances is the value you set here.
If multiple elastic policies are configured, the system calculates the Minimum Number Of Instances required when each policy is triggered. The system then uses the highest value among the active policies as the current Minimum Number Of Instances.
For more information, see How is the current minimum number of instances calculated?.
1
Concurrency Per Instance
You can configure multiple concurrent requests for a single GPU function instance. This means a single instance can process multiple requests simultaneously. For more information, see Configure concurrency per instance.
Provisioned instances
Configuration Item
Description
Example
Instance Type
Select Provisioned Instance. Instances are allocated to the function from a pre-purchased provisioned resource pool.
Provisioned instances are recommended for scenarios where predictable costs, low latency, and high resource utilization are important to ensure business stability.
Provisioned Instance
Provisioned Resource Pool
A provisioned resource pool is a pool of provisioned instances that can be allocated to the target function. If your provisioned resource pool has insufficient capacity, click Scale-out in the Actions column and follow the on-screen instructions to expand it. For more information, see Provisioned resource pools (subscription).
Provisioned Resource Pool: fc-pool-****
GPU Card Type: Ada
Specifications
Set the GPU Memory, vCPU, Memory, and Disk specifications for the function based on your business needs. After you set the specifications, the usage of each resource is calculated by multiplying the specification by the duration of use. For more information, see Billing overview.
NoteAll directories on the disk are writable. The disk space is shared.
The disk is tied to the instance lifetime of the underlying function. When the system reclaims the instance, data on the disk is lost. If you need persistent storage, you can mount a NAS file system or an OSS bucket. For more information, see Configure a NAS file system and Configure Object Storage Service.
GPU Memory: 48 GB
vCPU: 8 vCPU
Memory: 64 GB
Disk: 512 MB (not billed, Function Compute provides a free quota of 10 GB disk space)
Number Of Provisioned Instances
Allocate a number of provisioned instances to the target function based on the resources available in the provisioned resource pool.
1
Concurrency Per Instance
You can configure multiple concurrent requests for a single GPU function instance. This means a single instance can process multiple requests simultaneously. For more information, see Configure concurrency per instance.
20
Function Code: Configure the function's runtime environment and code.
Configuration Item
Description
Example
Runtime Environment
Use Sample Image: Select a sample image provided by Function Compute to quickly deploy an image-based function. Select the target image from the image list under the Container Image configuration item.
Use Image from ACR: Under the Container Image configuration item, click Select Image From ACR. In the Select Container Image panel, select the created Container Registry instance and ACR image repository. Then, find the target image in the image area below and click Select in the Actions column. For more information, see Create a function that uses a custom image.
Container Image
Select the target image.
SpringBoot Web Application Sample Image
Startup Command
The startup command for the program. If you do not configure a startup command, the Entrypoint/CMD from the image is used by default.
None
Listener Port
The port that the HTTP server in your code listens on.
9000
Execution Timeout
Set the timeout period. The default Execution Timeout is 60 seconds, and the maximum is 86400 seconds.
60
Instance Prefetch: In AI inference scenarios, you can configure instance prefetch to pre-warm the model. This eliminates the cold start latency for the first request.
Configuration Item
Description
Example
Instance Prefetch
Instance Prefetch
Configure an Initializer hook to pre-warm the instance and optimize cold starts. The hook runs a specified script or calls an interface to load the model after the function instance starts but before it processes requests.
For more information about Initializer hooks, see Configure the instance lifecycle.
Enabled
Timeout
Set the timeout period for the Initializer hook.
60
Prefetch Program Type
You can configure two types of Initializer hooks to pre-warm the model: Execute Instruction and Invoke Code.
Execute Instruction
Instruction Content
Configure the content of the instruction to execute. You can use custom shell implementations, such as
/bin/bash,/bin/sh,/bin/csh, and/bin/zsh. Make sure the function's runtime environment supports the selected shell.Use a role to grant Function Compute permissions to access other Alibaba Cloud services
mytestrole
fc.auto.create.vpc.1632317****
fc.auto.create.vswitch.vpc-bp1p8248****
fc.auto.create.SecurityGroup.vsw-bp15ftbbbbd****
ImportantUTC
key : value
{ "BUCKET_NAME": "MY_BUCKET", "TABLE_NAME": "MY_TABLE" }
Edit a function
After a function is created, you can change its image by editing the runtime on the Configuration tab of the function details page.
For information about other modifications, such as changing environment variables or log storage settings, see Configure a function.
References
Function Compute provides four function types for different scenarios: event functions, web functions, task functions, and GPU functions. For information about how to choose a function type for your scenario, see Technology selection guide.
In addition to the console, Function Compute provides APIs and the Serverless Devs tool to manage functions. For more information, see CreateFunction and Quick Start for Serverless Devs.
If a function execution times out, see What do I do if a "Function timeout" error occurs?.
Functions that are invoked infrequently may have longer invocation times. For more information about the reason, see Why do infrequently used functions have long invocation times?. To eliminate the impact of cold start latency, set the minimum number of instances to 1 or more.